Make Load3D model_file optional by adding "none" choice

.ci/windows_intel_base_files/run_intel_gpu.bat

@@ -0,0 +1,2 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+pause

blueprints/Brightness and Contrast.json

@@ -182,7 +182,7 @@
               ]
             },
             "widgets_values": [
-              50
+              0
             ]
           },
           {

blueprints/Glow.json

@@ -316,7 +316,7 @@
               "step": 1
             },
             "widgets_values": [
-              30
+              0
             ]
           },
           {

comfy/ldm/ernie/model.py

@@ -0,0 +1,303 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    assert dim % 2 == 0
+    if not comfy.model_management.supports_fp64(pos.device):
+        device = torch.device("cpu")
+    else:
+        device = pos.device
+
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=device) / dim
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos, omega)
+    out = torch.stack([torch.cos(out), torch.sin(out)], dim=0)
+    return out.to(dtype=torch.float32, device=pos.device)
+
+def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+    rot_dim = freqs_cis.shape[-1]
+    x, x_pass = x_in[..., :rot_dim], x_in[..., rot_dim:]
+    cos_ = freqs_cis[0]
+    sin_ = freqs_cis[1]
+    x1, x2 = x.chunk(2, dim=-1)
+    x_rotated = torch.cat((-x2, x1), dim=-1)
+    return torch.cat((x * cos_ + x_rotated * sin_, x_pass), dim=-1)
+
+class ErnieImageEmbedND3(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: tuple):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = list(axes_dim)
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        emb = torch.cat([rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(3)], dim=-1)
+        emb = emb.unsqueeze(3)  # [2, B, S, 1, head_dim//2]
+        return torch.stack([emb, emb], dim=-1).reshape(*emb.shape[:-1], -1)  # [B, S, 1, head_dim]
+
+class ErnieImagePatchEmbedDynamic(nn.Module):
+    def __init__(self, in_channels: int, embed_dim: int, patch_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.proj = operations.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        batch_size, dim, height, width = x.shape
+        return x.reshape(batch_size, dim, height * width).transpose(1, 2).contiguous()
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, flip_sin_to_cos: bool = False):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device) / half_dim
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+        if self.flip_sin_to_cos:
+            emb = torch.cat([torch.cos(emb), torch.sin(emb)], dim=-1)
+        else:
+            emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+        return emb
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.linear_1 = Linear(in_channels, time_embed_dim, bias=True, device=device, dtype=dtype)
+        self.act = nn.SiLU()
+        self.linear_2 = Linear(time_embed_dim, time_embed_dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+class ErnieImageAttention(nn.Module):
+    def __init__(self, query_dim: int, heads: int, dim_head: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim_head
+        self.inner_dim = heads * dim_head
+
+        Linear = operations.Linear
+        RMSNorm = operations.RMSNorm
+
+        self.to_q = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_k = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_v = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+
+        self.norm_q = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+
+        self.to_out = nn.ModuleList([Linear(self.inner_dim, query_dim, bias=False, device=device, dtype=dtype)])
+
+    def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None, image_rotary_emb: torch.Tensor = None) -> torch.Tensor:
+        B, S, _ = x.shape
+
+        q_flat = self.to_q(x)
+        k_flat = self.to_k(x)
+        v_flat = self.to_v(x)
+
+        query = q_flat.view(B, S, self.heads, self.head_dim)
+        key = k_flat.view(B, S, self.heads, self.head_dim)
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        query, key = query.to(x.dtype), key.to(x.dtype)
+
+        q_flat = query.reshape(B, S, -1)
+        k_flat = key.reshape(B, S, -1)
+
+        hidden_states = optimized_attention(q_flat, k_flat, v_flat, self.heads, mask=attention_mask)
+
+        return self.to_out[0](hidden_states)
+
+class ErnieImageFeedForward(nn.Module):
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.gate_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.linear_fc2 = Linear(ffn_hidden_size, hidden_size, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear_fc2(self.up_proj(x) * F.gelu(self.gate_proj(x)))
+
+class ErnieImageSharedAdaLNBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, ffn_hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        RMSNorm = operations.RMSNorm
+
+        self.adaLN_sa_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.self_attention = ErnieImageAttention(
+            query_dim=hidden_size,
+            dim_head=hidden_size // num_heads,
+            heads=num_heads,
+            eps=eps,
+            operations=operations,
+            device=device,
+            dtype=dtype
+        )
+        self.adaLN_mlp_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.mlp = ErnieImageFeedForward(hidden_size, ffn_hidden_size, operations=operations, device=device, dtype=dtype)
+
+    def forward(self, x, rotary_pos_emb, temb, attention_mask=None):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = temb
+
+        residual = x
+        x_norm = self.adaLN_sa_ln(x)
+        x_norm = (x_norm.float() * (1 + scale_msa.float()) + shift_msa.float()).to(x.dtype)
+
+        attn_out = self.self_attention(x_norm, attention_mask=attention_mask, image_rotary_emb=rotary_pos_emb)
+        x = residual + (gate_msa.float() * attn_out.float()).to(x.dtype)
+
+        residual = x
+        x_norm = self.adaLN_mlp_ln(x)
+        x_norm = (x_norm.float() * (1 + scale_mlp.float()) + shift_mlp.float()).to(x.dtype)
+
+        return residual + (gate_mlp.float() * self.mlp(x_norm).float()).to(x.dtype)
+
+class ErnieImageAdaLNContinuous(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        LayerNorm = operations.LayerNorm
+        Linear = operations.Linear
+        self.norm = LayerNorm(hidden_size, elementwise_affine=False, eps=eps, device=device, dtype=dtype)
+        self.linear = Linear(hidden_size, hidden_size * 2, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        scale, shift = self.linear(conditioning).chunk(2, dim=-1)
+        x = self.norm(x)
+        x = x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+        return x
+
+class ErnieImageModel(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        num_attention_heads: int = 32,
+        num_layers: int = 36,
+        ffn_hidden_size: int = 12288,
+        in_channels: int = 128,
+        out_channels: int = 128,
+        patch_size: int = 1,
+        text_in_dim: int = 3072,
+        rope_theta: int = 256,
+        rope_axes_dim: tuple = (32, 48, 48),
+        eps: float = 1e-6,
+        qk_layernorm: bool = True,
+        device=None,
+        dtype=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.hidden_size = hidden_size
+        self.num_heads = num_attention_heads
+        self.head_dim = hidden_size // num_attention_heads
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+
+        Linear = operations.Linear
+
+        self.x_embedder = ErnieImagePatchEmbedDynamic(in_channels, hidden_size, patch_size, operations, device, dtype)
+        self.text_proj = Linear(text_in_dim, hidden_size, bias=False, device=device, dtype=dtype) if text_in_dim != hidden_size else None
+
+        self.time_proj = Timesteps(hidden_size, flip_sin_to_cos=False)
+        self.time_embedding = TimestepEmbedding(hidden_size, hidden_size, operations, device, dtype)
+
+        self.pos_embed = ErnieImageEmbedND3(dim=self.head_dim, theta=rope_theta, axes_dim=rope_axes_dim)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            Linear(hidden_size, 6 * hidden_size, device=device, dtype=dtype)
+        )
+
+        self.layers = nn.ModuleList([
+            ErnieImageSharedAdaLNBlock(hidden_size, num_attention_heads, ffn_hidden_size, eps, operations, device, dtype)
+            for _ in range(num_layers)
+        ])
+
+        self.final_norm = ErnieImageAdaLNContinuous(hidden_size, eps, operations, device, dtype)
+        self.final_linear = Linear(hidden_size, patch_size * patch_size * out_channels, device=device, dtype=dtype)
+
+    def forward(self, x, timesteps, context, **kwargs):
+        device, dtype = x.device, x.dtype
+        B, C, H, W = x.shape
+        p, Hp, Wp = self.patch_size, H // self.patch_size, W // self.patch_size
+        N_img = Hp * Wp
+
+        img_bsh = self.x_embedder(x)
+
+        text_bth = context
+        if self.text_proj is not None and text_bth.numel() > 0:
+            text_bth = self.text_proj(text_bth)
+        Tmax = text_bth.shape[1]
+
+        hidden_states = torch.cat([img_bsh, text_bth], dim=1)
+
+        text_ids = torch.zeros((B, Tmax, 3), device=device, dtype=torch.float32)
+        text_ids[:, :, 0] = torch.linspace(0, Tmax - 1, steps=Tmax, device=x.device, dtype=torch.float32)
+        index = float(Tmax)
+
+        transformer_options = kwargs.get("transformer_options", {})
+        rope_options = transformer_options.get("rope_options", None)
+
+        h_len, w_len = float(Hp), float(Wp)
+        h_offset, w_offset = 0.0, 0.0
+
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        image_ids = torch.zeros((Hp, Wp, 3), device=device, dtype=torch.float32)
+        image_ids[:, :, 0] = image_ids[:, :, 1] + index
+        image_ids[:, :, 1] = image_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=Hp, device=device, dtype=torch.float32).unsqueeze(1)
+        image_ids[:, :, 2] = image_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=Wp, device=device, dtype=torch.float32).unsqueeze(0)
+
+        image_ids = image_ids.view(1, N_img, 3).expand(B, -1, -1)
+
+        rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1)).to(x.dtype)
+        del image_ids, text_ids
+
+        sample = self.time_proj(timesteps.to(dtype)).to(self.time_embedding.linear_1.weight.dtype)
+        c = self.time_embedding(sample)
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = [
+            t.unsqueeze(1).contiguous() for t in self.adaLN_modulation(c).chunk(6, dim=-1)
+        ]
+
+        temb = [shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp]
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, rotary_pos_emb, temb)
+
+        hidden_states = self.final_norm(hidden_states, c).type_as(hidden_states)
+
+        patches = self.final_linear(hidden_states)[:, :N_img, :]
+        output = (
+            patches.view(B, Hp, Wp, p, p, self.out_channels)
+            .permute(0, 5, 1, 3, 2, 4)
+            .contiguous()
+            .view(B, self.out_channels, H, W)
+        )
+
+        return output

comfy/ldm/flux/math.py

@@ -16,7 +16,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
+    if not comfy.model_management.supports_fp64(pos.device):
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/modules/sdpose.py

@@ -90,7 +90,7 @@ class HeatmapHead(torch.nn.Module):
                 origin_max = np.max(hm[k])
                 dr = np.zeros((H + 2 * border, W + 2 * border), dtype=np.float32)
                 dr[border:-border, border:-border] = hm[k].copy()
-                dr = gaussian_filter(dr, sigma=2.0)
+                dr = gaussian_filter(dr, sigma=2.0, truncate=2.5)
                 hm[k] = dr[border:-border, border:-border].copy()
                 cur_max = np.max(hm[k])
                 if cur_max > 0:

comfy/model_base.py

@@ -53,6 +53,7 @@ import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
 import comfy.ldm.ace.ace_step15
 import comfy.ldm.rt_detr.rtdetr_v4
+import comfy.ldm.ernie.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -1962,3 +1963,14 @@ class Kandinsky5Image(Kandinsky5):
 class RT_DETR_v4(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.rt_detr.rtdetr_v4.RTv4)
+
+class ErnieImage(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ernie.model.ErnieImageModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out

comfy/model_detection.py

@@ -713,6 +713,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["enc_h"] = state_dict['{}encoder.pan_blocks.1.cv4.conv.weight'.format(key_prefix)].shape[0]
         return dit_config
 
+    if '{}layers.0.mlp.linear_fc2.weight'.format(key_prefix) in state_dict_keys: # Ernie Image
+        dit_config = {}
+        dit_config["image_model"] = "ernie"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -1732,6 +1732,21 @@ def supports_mxfp8_compute(device=None):
 
     return True
 
+def supports_fp64(device=None):
+    if is_device_mps(device):
+        return False
+
+    if is_intel_xpu():
+        return False
+
+    if is_directml_enabled():
+        return False
+
+    if is_ixuca():
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):

comfy/sd.py

@@ -62,6 +62,7 @@ import comfy.text_encoders.anima
 import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
 import comfy.text_encoders.qwen35
+import comfy.text_encoders.ernie
 
 import comfy.model_patcher
 import comfy.lora
@@ -1235,6 +1236,7 @@ class TEModel(Enum):
     QWEN35_4B = 25
     QWEN35_9B = 26
     QWEN35_27B = 27
+    MINISTRAL_3_3B = 28
 
 
 def detect_te_model(sd):
@@ -1301,6 +1303,8 @@ def detect_te_model(sd):
                 return TEModel.MISTRAL3_24B
             else:
                 return TEModel.MISTRAL3_24B_PRUNED_FLUX2
+        if weight.shape[0] == 3072:
+            return TEModel.MINISTRAL_3_3B
 
         return TEModel.LLAMA3_8
     return None
@@ -1458,6 +1462,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif te_model == TEModel.QWEN3_06B:
             clip_target.clip = comfy.text_encoders.anima.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.anima.AnimaTokenizer
+        elif te_model == TEModel.MINISTRAL_3_3B:
+            clip_target.clip = comfy.text_encoders.ernie.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ernie.ErnieTokenizer
+            tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None)
         else:
             # clip_l
             if clip_type == CLIPType.SD3:

comfy/supported_models.py

@@ -26,6 +26,7 @@ import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
 import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
+import comfy.text_encoders.ernie
 
 from . import supported_models_base
 from . import latent_formats
@@ -1749,6 +1750,37 @@ class RT_DETR_v4(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return None
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, LongCatImage, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImagePixelSpace, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima, RT_DETR_v4]
+
+class ErnieImage(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "ernie",
+    }
+
+    sampling_settings = {
+        "multiplier": 1000.0,
+        "shift": 3.0,
+    }
+
+    memory_usage_factor = 10.0
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ErnieImage(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}ministral3_3b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.ernie.ErnieTokenizer, comfy.text_encoders.ernie.te(**hunyuan_detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, LongCatImage, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImagePixelSpace, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima, RT_DETR_v4, ErnieImage]
 
 models += [SVD_img2vid]

comfy/text_encoders/ernie.py

@@ -0,0 +1,38 @@
+from .flux import Mistral3Tokenizer
+from comfy import sd1_clip
+import comfy.text_encoders.llama
+
+class Ministral3_3BTokenizer(Mistral3Tokenizer):
+    def __init__(self, embedding_directory=None, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_data={}):
+        return super().__init__(embedding_directory=embedding_directory, embedding_size=embedding_size, embedding_key=embedding_key, tokenizer_data=tokenizer_data)
+
+class ErnieTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="ministral3_3b", tokenizer=Mistral3Tokenizer)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
+        tokens = super().tokenize_with_weights(text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        return tokens
+
+
+class Ministral3_3BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
+        textmodel_json_config = {}
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 1, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Ministral3_3B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class ErnieTEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, name="ministral3_3b", clip_model=Ministral3_3BModel):
+        super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
+
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class ErnieTEModel_(ErnieTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return ErnieTEModel

comfy/text_encoders/flux.py

@@ -116,9 +116,9 @@ class MistralTokenizerClass:
         return LlamaTokenizerFast(**kwargs)
 
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
+    def __init__(self, embedding_directory=None, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=embedding_size, embedding_key=embedding_key, tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, disable_weights=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}

comfy/text_encoders/llama.py

@@ -60,6 +60,29 @@ class Mistral3Small24BConfig:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Ministral3_3BConfig:
+    vocab_size: int = 131072
+    hidden_size: int = 3072
+    intermediate_size: int = 9216
+    num_hidden_layers: int = 26
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 262144
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = None
+    k_norm = None
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
 @dataclass
 class Qwen25_3BConfig:
     vocab_size: int = 151936
@@ -946,6 +969,15 @@ class Mistral3Small24B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Ministral3_3B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Ministral3_3BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen25_3B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy_api_nodes/nodes_grok.py

@@ -558,7 +558,7 @@ class GrokVideoReferenceNode(IO.ComfyNode):
                 (
                   $res := $lookup(widgets, "model.resolution");
                   $dur := $lookup(widgets, "model.duration");
-                  $refs := inputGroups["model.reference_images"];
+                  $refs := $lookup(inputGroups, "model.reference_images");
                   $rate := $res = "720p" ? 0.07 : 0.05;
                   $price := ($rate * $dur + 0.002 * $refs) * 1.43;
                   {"type":"usd","usd": $price}

comfy_extras/nodes_glsl.py

@@ -1,68 +1,85 @@
 import os
 import sys
 import re
-import ctypes
 import logging
+import ctypes.util
+import importlib.util
 from typing import TypedDict
 
 import numpy as np
 import torch
 
 import nodes
-import comfy_angle
 from comfy_api.latest import ComfyExtension, io, ui
 from typing_extensions import override
+from utils.install_util import get_missing_requirements_message
 
 logger = logging.getLogger(__name__)
 
 
-def _preload_angle():
-    egl_path = comfy_angle.get_egl_path()
-    gles_path = comfy_angle.get_glesv2_path()
+def _check_opengl_availability():
+    """Early check for OpenGL availability. Raises RuntimeError if unlikely to work."""
+    logger.debug("_check_opengl_availability: starting")
+    missing = []
 
-    if sys.platform == "win32":
-        angle_dir = comfy_angle.get_lib_dir()
-        os.add_dll_directory(angle_dir)
-        os.environ["PATH"] = angle_dir + os.pathsep + os.environ.get("PATH", "")
+    # Check Python packages (using find_spec to avoid importing)
+    logger.debug("_check_opengl_availability: checking for glfw package")
+    if importlib.util.find_spec("glfw") is None:
+        missing.append("glfw")
 
-    mode = 0 if sys.platform == "win32" else ctypes.RTLD_GLOBAL
-    ctypes.CDLL(str(egl_path), mode=mode)
-    ctypes.CDLL(str(gles_path), mode=mode)
+    logger.debug("_check_opengl_availability: checking for OpenGL package")
+    if importlib.util.find_spec("OpenGL") is None:
+        missing.append("PyOpenGL")
+
+    if missing:
+        raise RuntimeError(
+            f"OpenGL dependencies not available.\n{get_missing_requirements_message()}\n"
+        )
+
+    # On Linux without display, check if headless backends are available
+    logger.debug(f"_check_opengl_availability: platform={sys.platform}")
+    if sys.platform.startswith("linux"):
+        has_display = os.environ.get("DISPLAY") or os.environ.get("WAYLAND_DISPLAY")
+        logger.debug(f"_check_opengl_availability: has_display={bool(has_display)}")
+        if not has_display:
+            # Check for EGL or OSMesa libraries
+            logger.debug("_check_opengl_availability: checking for EGL library")
+            has_egl = ctypes.util.find_library("EGL")
+            logger.debug("_check_opengl_availability: checking for OSMesa library")
+            has_osmesa = ctypes.util.find_library("OSMesa")
+
+            # Error disabled for CI as it fails this check
+            # if not has_egl and not has_osmesa:
+            #     raise RuntimeError(
+            #         "GLSL Shader node: No display and no headless backend (EGL/OSMesa) found.\n"
+            #         "See error below for installation instructions."
+            #     )
+            logger.debug(f"Headless mode: EGL={'yes' if has_egl else 'no'}, OSMesa={'yes' if has_osmesa else 'no'}")
+
+    logger.debug("_check_opengl_availability: completed")
 
 
-# Pre-load ANGLE *before* any PyOpenGL import so that the EGL platform
-# plugin picks up ANGLE's libEGL / libGLESv2 instead of system libs.
-_preload_angle()
-os.environ.setdefault("PYOPENGL_PLATFORM", "egl")
+# Run early check at import time
+logger.debug("nodes_glsl: running _check_opengl_availability at import time")
+_check_opengl_availability()
+
+# OpenGL modules - initialized lazily when context is created
+gl = None
+glfw = None
+EGL = None
 
 
-import OpenGL
-OpenGL.USE_ACCELERATE = False
+def _import_opengl():
+    """Import OpenGL module. Called after context is created."""
+    global gl
+    if gl is None:
+        logger.debug("_import_opengl: importing OpenGL.GL")
+        import OpenGL.GL as _gl
+        gl = _gl
+        logger.debug("_import_opengl: import completed")
+    return gl
 
 
-def _patch_find_library():
-    """PyOpenGL's EGL platform looks for 'EGL' and 'GLESv2' by short name
-    via ctypes.util.find_library, but ANGLE ships as 'libEGL' and
-    'libGLESv2'.  Patch find_library to return the full ANGLE paths so
-    PyOpenGL loads the same libraries we pre-loaded."""
-    if sys.platform == "linux":
-        return
-    import ctypes.util
-    _orig = ctypes.util.find_library
-    def _patched(name):
-        if name == 'EGL':
-            return comfy_angle.get_egl_path()
-        if name == 'GLESv2':
-            return comfy_angle.get_glesv2_path()
-        return _orig(name)
-    ctypes.util.find_library = _patched
-
-
-_patch_find_library()
-
-from OpenGL import EGL
-from OpenGL import GLES3 as gl
-
 class SizeModeInput(TypedDict):
     size_mode: str
     width: int
@@ -85,7 +102,7 @@ MAX_OUTPUTS = 4     # fragColor0-3 (MRT)
 # (-1,-1)---(3,-1)
 #
 # v_texCoord is computed from clip space: * 0.5 + 0.5 maps (-1,1) -> (0,1)
-VERTEX_SHADER = """#version 300 es
+VERTEX_SHADER = """#version 330 core
 out vec2 v_texCoord;
 void main() {
     vec2 verts[3] = vec2[](vec2(-1, -1), vec2(3, -1), vec2(-1, 3));
@@ -109,99 +126,14 @@ void main() {
 """
 
 
-
-def _egl_attribs(*values):
-    """Build an EGL_NONE-terminated EGLint attribute array."""
-    vals = list(values) + [EGL.EGL_NONE]
-    return (ctypes.c_int32 * len(vals))(*vals)
-
-
-# EGL platform extension constants
-EGL_PLATFORM_ANGLE_ANGLE = 0x3202
-EGL_PLATFORM_ANGLE_TYPE_ANGLE = 0x3203
-EGL_PLATFORM_ANGLE_TYPE_VULKAN_ANGLE = 0x3450
-EGL_MESA_PLATFORM_SURFACELESS = 0x31DD
-
-
-_eglGetPlatformDisplayEXT = None
-
-def _get_egl_platform_display_ext(platform, native_display, attribs):
-    """Call eglGetPlatformDisplayEXT via ctypes (extension, not in PyOpenGL)."""
-    global _eglGetPlatformDisplayEXT
-    if _eglGetPlatformDisplayEXT is None:
-        from OpenGL import platform as _plat
-        egl_lib = _plat.PLATFORM.EGL
-        _get_proc = egl_lib.eglGetProcAddress
-        _get_proc.restype = ctypes.c_void_p
-        _get_proc.argtypes = [ctypes.c_char_p]
-        ptr = _get_proc(b"eglGetPlatformDisplayEXT")
-        if not ptr:
-            return None
-        func_type = ctypes.CFUNCTYPE(ctypes.c_void_p, ctypes.c_uint32, ctypes.c_void_p, ctypes.c_void_p)
-        _eglGetPlatformDisplayEXT = func_type(ptr)
-
-    raw = _eglGetPlatformDisplayEXT(platform, native_display, attribs)
-    if not raw:
-        return None
-    return ctypes.cast(raw, EGL.EGLDisplay)
-
-
-def _get_egl_display():
-    """Get an EGL display, trying the default first then ANGLE's Vulkan
-    platform for headless environments without a display server."""
-    failures = []
-
-    # Try the default display first (works when X11/Wayland is available)
-    display = EGL.eglGetDisplay(EGL.EGL_DEFAULT_DISPLAY)
-    if display:
-        major, minor = ctypes.c_int32(0), ctypes.c_int32(0)
-        try:
-            if EGL.eglInitialize(display, ctypes.byref(major), ctypes.byref(minor)):
-                return display, major.value, minor.value
-        except Exception as e:
-            failures.append(f"default: {e}")
-
-    logger.info("Default EGL display unavailable, trying headless fallbacks")
-
-    # Headless fallback strategies, tried in order:
-    headless_strategies = [
-        ("surfaceless", EGL_MESA_PLATFORM_SURFACELESS, None, None),
-        ("ANGLE Vulkan", EGL_PLATFORM_ANGLE_ANGLE, None,
-         _egl_attribs(EGL_PLATFORM_ANGLE_TYPE_ANGLE, EGL_PLATFORM_ANGLE_TYPE_VULKAN_ANGLE)),
-    ]
-
-    for name, platform, native_display, attribs in headless_strategies:
-        display = _get_egl_platform_display_ext(platform, native_display, attribs)
-        if not display:
-            failures.append(f"{name}: eglGetPlatformDisplayEXT returned no display")
-            continue
-        major, minor = ctypes.c_int32(0), ctypes.c_int32(0)
-        try:
-            if EGL.eglInitialize(display, ctypes.byref(major), ctypes.byref(minor)):
-                logger.info(f"Using EGL {name} platform (headless)")
-                return display, major.value, minor.value
-            failures.append(f"{name}: eglInitialize returned false")
-        except Exception as e:
-            failures.append(f"{name}: {e}")
-            continue
-
-    details = "\n".join(f"  - {f}" for f in failures)
-    raise RuntimeError(
-        "Failed to initialize EGL display.\n"
-        "No display server and no headless EGL platform available.\n"
-        f"Tried:\n{details}\n"
-        "Ensure GPU drivers are installed or set DISPLAY for a virtual framebuffer."
-    )
-
-
-def _gl_str(name):
-    """Get an OpenGL string parameter."""
-    v = gl.glGetString(name)
-    if not v:
-        return "Unknown"
-    if isinstance(v, bytes):
-        return v.decode(errors="replace")
-    return ctypes.string_at(v).decode(errors="replace")
+def _convert_es_to_desktop(source: str) -> str:
+    """Convert GLSL ES (WebGL) shader source to desktop GLSL 330 core."""
+    # Remove any existing #version directive
+    source = re.sub(r"#version\s+\d+(\s+es)?\s*\n?", "", source, flags=re.IGNORECASE)
+    # Remove precision qualifiers (not needed in desktop GLSL)
+    source = re.sub(r"precision\s+(lowp|mediump|highp)\s+\w+\s*;\s*\n?", "", source)
+    # Prepend desktop GLSL version
+    return "#version 330 core\n" + source
 
 
 def _detect_output_count(source: str) -> int:
@@ -227,8 +159,163 @@ def _detect_pass_count(source: str) -> int:
     return 1
 
 
+def _init_glfw():
+    """Initialize GLFW. Returns (window, glfw_module). Raises RuntimeError on failure."""
+    logger.debug("_init_glfw: starting")
+    # On macOS, glfw.init() must be called from main thread or it hangs forever
+    if sys.platform == "darwin":
+        logger.debug("_init_glfw: skipping on macOS")
+        raise RuntimeError("GLFW backend not supported on macOS")
+
+    logger.debug("_init_glfw: importing glfw module")
+    import glfw as _glfw
+
+    logger.debug("_init_glfw: calling glfw.init()")
+    if not _glfw.init():
+        raise RuntimeError("glfw.init() failed")
+
+    try:
+        logger.debug("_init_glfw: setting window hints")
+        _glfw.window_hint(_glfw.VISIBLE, _glfw.FALSE)
+        _glfw.window_hint(_glfw.CONTEXT_VERSION_MAJOR, 3)
+        _glfw.window_hint(_glfw.CONTEXT_VERSION_MINOR, 3)
+        _glfw.window_hint(_glfw.OPENGL_PROFILE, _glfw.OPENGL_CORE_PROFILE)
+
+        logger.debug("_init_glfw: calling create_window()")
+        window = _glfw.create_window(64, 64, "ComfyUI GLSL", None, None)
+        if not window:
+            raise RuntimeError("glfw.create_window() failed")
+
+        logger.debug("_init_glfw: calling make_context_current()")
+        _glfw.make_context_current(window)
+        logger.debug("_init_glfw: completed successfully")
+        return window, _glfw
+    except Exception:
+        logger.debug("_init_glfw: failed, terminating glfw")
+        _glfw.terminate()
+        raise
+
+
+def _init_egl():
+    """Initialize EGL for headless rendering. Returns (display, context, surface, EGL_module). Raises RuntimeError on failure."""
+    logger.debug("_init_egl: starting")
+    from OpenGL import EGL as _EGL
+    from OpenGL.EGL import (
+        eglGetDisplay, eglInitialize, eglChooseConfig, eglCreateContext,
+        eglMakeCurrent, eglCreatePbufferSurface, eglBindAPI,
+        eglTerminate, eglDestroyContext, eglDestroySurface,
+        EGL_DEFAULT_DISPLAY, EGL_NO_CONTEXT, EGL_NONE,
+        EGL_SURFACE_TYPE, EGL_PBUFFER_BIT, EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+        EGL_RED_SIZE, EGL_GREEN_SIZE, EGL_BLUE_SIZE, EGL_ALPHA_SIZE, EGL_DEPTH_SIZE,
+        EGL_WIDTH, EGL_HEIGHT, EGL_OPENGL_API,
+    )
+    logger.debug("_init_egl: imports completed")
+
+    display = None
+    context = None
+    surface = None
+
+    try:
+        logger.debug("_init_egl: calling eglGetDisplay()")
+        display = eglGetDisplay(EGL_DEFAULT_DISPLAY)
+        if display == _EGL.EGL_NO_DISPLAY:
+            raise RuntimeError("eglGetDisplay() failed")
+
+        logger.debug("_init_egl: calling eglInitialize()")
+        major, minor = _EGL.EGLint(), _EGL.EGLint()
+        if not eglInitialize(display, major, minor):
+            display = None  # Not initialized, don't terminate
+            raise RuntimeError("eglInitialize() failed")
+        logger.debug(f"_init_egl: EGL version {major.value}.{minor.value}")
+
+        config_attribs = [
+            EGL_SURFACE_TYPE, EGL_PBUFFER_BIT,
+            EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+            EGL_RED_SIZE, 8, EGL_GREEN_SIZE, 8, EGL_BLUE_SIZE, 8, EGL_ALPHA_SIZE, 8,
+            EGL_DEPTH_SIZE, 0, EGL_NONE
+        ]
+        configs = (_EGL.EGLConfig * 1)()
+        num_configs = _EGL.EGLint()
+        if not eglChooseConfig(display, config_attribs, configs, 1, num_configs) or num_configs.value == 0:
+            raise RuntimeError("eglChooseConfig() failed")
+        config = configs[0]
+        logger.debug(f"_init_egl: config chosen, num_configs={num_configs.value}")
+
+        if not eglBindAPI(EGL_OPENGL_API):
+            raise RuntimeError("eglBindAPI() failed")
+
+        logger.debug("_init_egl: calling eglCreateContext()")
+        context_attribs = [
+            _EGL.EGL_CONTEXT_MAJOR_VERSION, 3,
+            _EGL.EGL_CONTEXT_MINOR_VERSION, 3,
+            _EGL.EGL_CONTEXT_OPENGL_PROFILE_MASK, _EGL.EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT,
+            EGL_NONE
+        ]
+        context = eglCreateContext(display, config, EGL_NO_CONTEXT, context_attribs)
+        if context == EGL_NO_CONTEXT:
+            raise RuntimeError("eglCreateContext() failed")
+
+        logger.debug("_init_egl: calling eglCreatePbufferSurface()")
+        pbuffer_attribs = [EGL_WIDTH, 64, EGL_HEIGHT, 64, EGL_NONE]
+        surface = eglCreatePbufferSurface(display, config, pbuffer_attribs)
+        if surface == _EGL.EGL_NO_SURFACE:
+            raise RuntimeError("eglCreatePbufferSurface() failed")
+
+        logger.debug("_init_egl: calling eglMakeCurrent()")
+        if not eglMakeCurrent(display, surface, surface, context):
+            raise RuntimeError("eglMakeCurrent() failed")
+
+        logger.debug("_init_egl: completed successfully")
+        return display, context, surface, _EGL
+
+    except Exception:
+        logger.debug("_init_egl: failed, cleaning up")
+        # Clean up any resources on failure
+        if surface is not None:
+            eglDestroySurface(display, surface)
+        if context is not None:
+            eglDestroyContext(display, context)
+        if display is not None:
+            eglTerminate(display)
+        raise
+
+
+def _init_osmesa():
+    """Initialize OSMesa for software rendering. Returns (context, buffer). Raises RuntimeError on failure."""
+    import ctypes
+
+    logger.debug("_init_osmesa: starting")
+    os.environ["PYOPENGL_PLATFORM"] = "osmesa"
+
+    logger.debug("_init_osmesa: importing OpenGL.osmesa")
+    from OpenGL import GL as _gl
+    from OpenGL.osmesa import (
+        OSMesaCreateContextExt, OSMesaMakeCurrent, OSMesaDestroyContext,
+        OSMESA_RGBA,
+    )
+    logger.debug("_init_osmesa: imports completed")
+
+    ctx = OSMesaCreateContextExt(OSMESA_RGBA, 24, 0, 0, None)
+    if not ctx:
+        raise RuntimeError("OSMesaCreateContextExt() failed")
+
+    width, height = 64, 64
+    buffer = (ctypes.c_ubyte * (width * height * 4))()
+
+    logger.debug("_init_osmesa: calling OSMesaMakeCurrent()")
+    if not OSMesaMakeCurrent(ctx, buffer, _gl.GL_UNSIGNED_BYTE, width, height):
+        OSMesaDestroyContext(ctx)
+        raise RuntimeError("OSMesaMakeCurrent() failed")
+
+    logger.debug("_init_osmesa: completed successfully")
+    return ctx, buffer
+
+
 class GLContext:
-    """Manages an OpenGL ES 3.0 context via EGL/ANGLE (singleton)."""
+    """Manages OpenGL context and resources for shader execution.
+
+    Tries backends in order: GLFW (desktop) → EGL (headless GPU) → OSMesa (software).
+    """
 
     _instance = None
     _initialized = False
@@ -240,105 +327,131 @@ class GLContext:
 
     def __init__(self):
         if GLContext._initialized:
+            logger.debug("GLContext.__init__: already initialized, skipping")
             return
 
+        logger.debug("GLContext.__init__: starting initialization")
+
+        global glfw, EGL
+
         import time
         start = time.perf_counter()
 
-        self._display = None
-        self._surface = None
-        self._context = None
+        self._backend = None
+        self._window = None
+        self._egl_display = None
+        self._egl_context = None
+        self._egl_surface = None
+        self._osmesa_ctx = None
+        self._osmesa_buffer = None
         self._vao = None
 
+        # Try backends in order: GLFW → EGL → OSMesa
+        errors = []
+
+        logger.debug("GLContext.__init__: trying GLFW backend")
         try:
-            self._display, self._egl_major, self._egl_minor = _get_egl_display()
+            self._window, glfw = _init_glfw()
+            self._backend = "glfw"
+            logger.debug("GLContext.__init__: GLFW backend succeeded")
+        except Exception as e:
+            logger.debug(f"GLContext.__init__: GLFW backend failed: {e}")
+            errors.append(("GLFW", e))
 
-            if not EGL.eglBindAPI(EGL.EGL_OPENGL_ES_API):
-                raise RuntimeError("eglBindAPI(EGL_OPENGL_ES_API) failed")
+        if self._backend is None:
+            logger.debug("GLContext.__init__: trying EGL backend")
+            try:
+                self._egl_display, self._egl_context, self._egl_surface, EGL = _init_egl()
+                self._backend = "egl"
+                logger.debug("GLContext.__init__: EGL backend succeeded")
+            except Exception as e:
+                logger.debug(f"GLContext.__init__: EGL backend failed: {e}")
+                errors.append(("EGL", e))
 
-            config = EGL.EGLConfig()
-            n_configs = ctypes.c_int32(0)
-            if not EGL.eglChooseConfig(
-                self._display,
-                _egl_attribs(
-                    EGL.EGL_RENDERABLE_TYPE, EGL.EGL_OPENGL_ES3_BIT,
-                    EGL.EGL_SURFACE_TYPE, EGL.EGL_PBUFFER_BIT,
-                    EGL.EGL_RED_SIZE, 8, EGL.EGL_GREEN_SIZE, 8,
-                    EGL.EGL_BLUE_SIZE, 8, EGL.EGL_ALPHA_SIZE, 8,
-                ),
-                ctypes.byref(config), 1, ctypes.byref(n_configs),
-            ) or n_configs.value == 0:
-                raise RuntimeError("eglChooseConfig() failed")
+        if self._backend is None:
+            logger.debug("GLContext.__init__: trying OSMesa backend")
+            try:
+                self._osmesa_ctx, self._osmesa_buffer = _init_osmesa()
+                self._backend = "osmesa"
+                logger.debug("GLContext.__init__: OSMesa backend succeeded")
+            except Exception as e:
+                logger.debug(f"GLContext.__init__: OSMesa backend failed: {e}")
+                errors.append(("OSMesa", e))
 
-            self._surface = EGL.eglCreatePbufferSurface(
-                self._display, config,
-                _egl_attribs(EGL.EGL_WIDTH, 64, EGL.EGL_HEIGHT, 64),
+        if self._backend is None:
+            if sys.platform == "win32":
+                platform_help = (
+                    "Windows: Ensure GPU drivers are installed and display is available.\n"
+                    "         CPU-only/headless mode is not supported on Windows."
+                )
+            elif sys.platform == "darwin":
+                platform_help = (
+                    "macOS: GLFW is not supported.\n"
+                    "  Install OSMesa via Homebrew: brew install mesa\n"
+                    "  Then: pip install PyOpenGL PyOpenGL-accelerate"
+                )
+            else:
+                platform_help = (
+                    "Linux: Install one of these backends:\n"
+                    "  Desktop:           sudo apt install libgl1-mesa-glx libglfw3\n"
+                    "  Headless with GPU: sudo apt install libegl1-mesa libgl1-mesa-dri\n"
+                    "  Headless (CPU):    sudo apt install libosmesa6"
+                )
+
+            error_details = "\n".join(f"  {name}: {err}" for name, err in errors)
+            raise RuntimeError(
+                f"Failed to create OpenGL context.\n\n"
+                f"Backend errors:\n{error_details}\n\n"
+                f"{platform_help}"
             )
-            if not self._surface:
-                raise RuntimeError("eglCreatePbufferSurface() failed")
 
-            self._context = EGL.eglCreateContext(
-                self._display, config, EGL.EGL_NO_CONTEXT,
-                _egl_attribs(EGL.EGL_CONTEXT_CLIENT_VERSION, 3),
-            )
-            if not self._context:
-                raise RuntimeError("eglCreateContext() failed")
+        # Now import OpenGL.GL (after context is current)
+        logger.debug("GLContext.__init__: importing OpenGL.GL")
+        _import_opengl()
 
-            if not EGL.eglMakeCurrent(self._display, self._surface, self._surface, self._context):
-                raise RuntimeError("eglMakeCurrent() failed")
-
-            self._vao = gl.glGenVertexArrays(1)
-            gl.glBindVertexArray(self._vao)
-
-        except Exception:
-            self._cleanup()
-            raise
+        # Create VAO (required for core profile, but OSMesa may use compat profile)
+        logger.debug("GLContext.__init__: creating VAO")
+        try:
+            vao = gl.glGenVertexArrays(1)
+            gl.glBindVertexArray(vao)
+            self._vao = vao  # Only store after successful bind
+            logger.debug("GLContext.__init__: VAO created successfully")
+        except Exception as e:
+            logger.debug(f"GLContext.__init__: VAO creation failed (may be expected for OSMesa): {e}")
+            # OSMesa with older Mesa may not support VAOs
+            # Clean up if we created but couldn't bind
+            if vao:
+                try:
+                    gl.glDeleteVertexArrays(1, [vao])
+                except Exception:
+                    pass
 
         elapsed = (time.perf_counter() - start) * 1000
 
-        renderer = _gl_str(gl.GL_RENDERER)
-        vendor = _gl_str(gl.GL_VENDOR)
-        version = _gl_str(gl.GL_VERSION)
+        # Log device info
+        renderer = gl.glGetString(gl.GL_RENDERER)
+        vendor = gl.glGetString(gl.GL_VENDOR)
+        version = gl.glGetString(gl.GL_VERSION)
+        renderer = renderer.decode() if renderer else "Unknown"
+        vendor = vendor.decode() if vendor else "Unknown"
+        version = version.decode() if version else "Unknown"
 
         GLContext._initialized = True
-        logger.info(f"GLSL context initialized in {elapsed:.1f}ms - EGL {self._egl_major}.{self._egl_minor}, {renderer} ({vendor}), GL {version}")
+        logger.info(f"GLSL context initialized in {elapsed:.1f}ms ({self._backend}) - {renderer} ({vendor}), GL {version}")
 
     def make_current(self):
-        if not EGL.eglMakeCurrent(self._display, self._surface, self._surface, self._context):
-            err = EGL.eglGetError()
-            raise RuntimeError(f"eglMakeCurrent() failed (EGL error: 0x{err:04X})")
+        if self._backend == "glfw":
+            glfw.make_context_current(self._window)
+        elif self._backend == "egl":
+            from OpenGL.EGL import eglMakeCurrent
+            eglMakeCurrent(self._egl_display, self._egl_surface, self._egl_surface, self._egl_context)
+        elif self._backend == "osmesa":
+            from OpenGL.osmesa import OSMesaMakeCurrent
+            OSMesaMakeCurrent(self._osmesa_ctx, self._osmesa_buffer, gl.GL_UNSIGNED_BYTE, 64, 64)
+
         if self._vao is not None:
             gl.glBindVertexArray(self._vao)
 
-    def _cleanup(self):
-        if not self._display:
-            return
-        try:
-            if self._vao is not None:
-                gl.glDeleteVertexArrays(1, [self._vao])
-                self._vao = None
-        except Exception:
-            pass
-        try:
-            EGL.eglMakeCurrent(self._display, EGL.EGL_NO_SURFACE, EGL.EGL_NO_SURFACE, EGL.EGL_NO_CONTEXT)
-        except Exception:
-            pass
-        try:
-            if self._context:
-                EGL.eglDestroyContext(self._display, self._context)
-        except Exception:
-            pass
-        try:
-            if self._surface:
-                EGL.eglDestroySurface(self._display, self._surface)
-        except Exception:
-            pass
-        try:
-            EGL.eglTerminate(self._display)
-        except Exception:
-            pass
-        self._display = None
-
 
 def _compile_shader(source: str, shader_type: int) -> int:
     """Compile a shader and return its ID."""
@@ -346,10 +459,8 @@ def _compile_shader(source: str, shader_type: int) -> int:
     gl.glShaderSource(shader, source)
     gl.glCompileShader(shader)
 
-    if not gl.glGetShaderiv(shader, gl.GL_COMPILE_STATUS):
-        error = gl.glGetShaderInfoLog(shader)
-        if isinstance(error, bytes):
-            error = error.decode(errors="replace")
+    if gl.glGetShaderiv(shader, gl.GL_COMPILE_STATUS) != gl.GL_TRUE:
+        error = gl.glGetShaderInfoLog(shader).decode()
         gl.glDeleteShader(shader)
         raise RuntimeError(f"Shader compilation failed:\n{error}")
 
@@ -373,10 +484,8 @@ def _create_program(vertex_source: str, fragment_source: str) -> int:
     gl.glDeleteShader(vertex_shader)
     gl.glDeleteShader(fragment_shader)
 
-    if not gl.glGetProgramiv(program, gl.GL_LINK_STATUS):
-        error = gl.glGetProgramInfoLog(program)
-        if isinstance(error, bytes):
-            error = error.decode(errors="replace")
+    if gl.glGetProgramiv(program, gl.GL_LINK_STATUS) != gl.GL_TRUE:
+        error = gl.glGetProgramInfoLog(program).decode()
         gl.glDeleteProgram(program)
         raise RuntimeError(f"Program linking failed:\n{error}")
 
@@ -421,6 +530,9 @@ def _render_shader_batch(
     ctx = GLContext()
     ctx.make_current()
 
+    # Convert from GLSL ES to desktop GLSL 330
+    fragment_source = _convert_es_to_desktop(fragment_code)
+
     # Detect how many outputs the shader actually uses
     num_outputs = _detect_output_count(fragment_code)
 
@@ -446,9 +558,9 @@ def _render_shader_batch(
     try:
         # Compile shaders (once for all batches)
         try:
-            program = _create_program(VERTEX_SHADER, fragment_code)
+            program = _create_program(VERTEX_SHADER, fragment_source)
         except RuntimeError:
-            logger.error(f"Fragment shader:\n{fragment_code}")
+            logger.error(f"Fragment shader:\n{fragment_source}")
             raise
 
         gl.glUseProgram(program)
@@ -611,13 +723,13 @@ def _render_shader_batch(
                     gl.glDrawArrays(gl.GL_TRIANGLES, 0, 3)
 
             # Read back outputs for this batch
-            gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, fbo)
+            # (glGetTexImage is synchronous, implicitly waits for rendering)
             batch_outputs = []
-            for i in range(num_outputs):
-                gl.glReadBuffer(gl.GL_COLOR_ATTACHMENT0 + i)
-                buf = np.empty((height, width, 4), dtype=np.float32)
-                gl.glReadPixels(0, 0, width, height, gl.GL_RGBA, gl.GL_FLOAT, buf)
-                batch_outputs.append(buf[::-1, :, :].copy())
+            for tex in output_textures:
+                gl.glBindTexture(gl.GL_TEXTURE_2D, tex)
+                data = gl.glGetTexImage(gl.GL_TEXTURE_2D, 0, gl.GL_RGBA, gl.GL_FLOAT)
+                img = np.frombuffer(data, dtype=np.float32).reshape(height, width, 4)
+                batch_outputs.append(img[::-1, :, :].copy())
 
             # Pad with black images for unused outputs
             black_img = np.zeros((height, width, 4), dtype=np.float32)
@@ -638,18 +750,18 @@ def _render_shader_batch(
         gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, 0)
         gl.glUseProgram(0)
 
-        if input_textures:
-            gl.glDeleteTextures(len(input_textures), input_textures)
-        if curve_textures:
-            gl.glDeleteTextures(len(curve_textures), curve_textures)
-        if output_textures:
-            gl.glDeleteTextures(len(output_textures), output_textures)
-        if ping_pong_textures:
-            gl.glDeleteTextures(len(ping_pong_textures), ping_pong_textures)
+        for tex in input_textures:
+            gl.glDeleteTextures(int(tex))
+        for tex in curve_textures:
+            gl.glDeleteTextures(int(tex))
+        for tex in output_textures:
+            gl.glDeleteTextures(int(tex))
+        for tex in ping_pong_textures:
+            gl.glDeleteTextures(int(tex))
         if fbo is not None:
             gl.glDeleteFramebuffers(1, [fbo])
-        if ping_pong_fbos:
-            gl.glDeleteFramebuffers(len(ping_pong_fbos), ping_pong_fbos)
+        for pp_fbo in ping_pong_fbos:
+            gl.glDeleteFramebuffers(1, [pp_fbo])
         if program is not None:
             gl.glDeleteProgram(program)
 

comfy_extras/nodes_load_3d.py

@@ -34,7 +34,7 @@ class Load3D(IO.ComfyNode):
             essentials_category="Basics",
             is_experimental=True,
             inputs=[
-                IO.Combo.Input("model_file", options=sorted(files), upload=IO.UploadType.model),
+                IO.Combo.Input("model_file", options=["none"] + sorted(files), upload=IO.UploadType.model),
                 IO.Load3D.Input("image"),
                 IO.Int.Input("width", default=1024, min=1, max=4096, step=1),
                 IO.Int.Input("height", default=1024, min=1, max=4096, step=1),
@@ -68,8 +68,12 @@ class Load3D(IO.ComfyNode):
 
             video = InputImpl.VideoFromFile(recording_video_path)
 
-        file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
-        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d)
+        file_3d = None
+        mesh_path = ""
+        if model_file and model_file != "none":
+            file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
+            mesh_path = model_file
+        return IO.NodeOutput(output_image, output_mask, mesh_path, normal_image, image['camera_info'], video, file_3d)
 
     process = execute  # TODO: remove
 

comfy_extras/nodes_rtdetr.py

@@ -32,10 +32,12 @@ class RTDETR_detect(io.ComfyNode):
     def execute(cls, model, image, threshold, class_name, max_detections) -> io.NodeOutput:
         B, H, W, C = image.shape
 
-        image_in = comfy.utils.common_upscale(image.movedim(-1, 1), 640, 640, "bilinear", crop="disabled")
-
         comfy.model_management.load_model_gpu(model)
-        results = model.model.diffusion_model(image_in, (W, H))  # list of B dicts
+        results = []
+        for i in range(0, B, 32):
+            batch = image[i:i + 32]
+            image_in = comfy.utils.common_upscale(batch.movedim(-1, 1), 640, 640, "bilinear", crop="disabled")
+            results.extend(model.model.diffusion_model(image_in, (W, H)))
 
         all_bbox_dicts = []
 

comfy_extras/nodes_sdpose.py

@@ -1,5 +1,6 @@
 import torch
 import comfy.utils
+import comfy.model_management
 import numpy as np
 import math
 import colorsys
@@ -410,7 +411,9 @@ class SDPoseDrawKeypoints(io.ComfyNode):
             pose_outputs.append(canvas)
 
         pose_outputs_np = np.stack(pose_outputs) if len(pose_outputs) > 1 else np.expand_dims(pose_outputs[0], 0)
-        final_pose_output = torch.from_numpy(pose_outputs_np).float() / 255.0
+        final_pose_output = torch.from_numpy(pose_outputs_np).to(
+            device=comfy.model_management.intermediate_device(),
+            dtype=comfy.model_management.intermediate_dtype()) / 255.0
         return io.NodeOutput(final_pose_output)
 
 class SDPoseKeypointExtractor(io.ComfyNode):
@@ -459,6 +462,27 @@ class SDPoseKeypointExtractor(io.ComfyNode):
         model_h = int(head.heatmap_size[0]) * 4   # e.g. 192 * 4 = 768
         model_w = int(head.heatmap_size[1]) * 4   # e.g. 256 * 4 = 1024
 
+        def _resize_to_model(imgs):
+            """Aspect-preserving resize + zero-pad BHWC images to (model_h, model_w). Returns (resized_bhwc, scale, pad_top, pad_left)."""
+            h, w = imgs.shape[-3], imgs.shape[-2]
+            scale = min(model_h / h, model_w / w)
+            sh, sw = int(round(h * scale)), int(round(w * scale))
+            pt, pl = (model_h - sh) // 2, (model_w - sw) // 2
+            chw = imgs.permute(0, 3, 1, 2).float()
+            scaled = comfy.utils.common_upscale(chw, sw, sh, upscale_method="bilinear", crop="disabled")
+            padded = torch.zeros(scaled.shape[0], scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
+            padded[:, :, pt:pt + sh, pl:pl + sw] = scaled
+            return padded.permute(0, 2, 3, 1), scale, pt, pl
+
+        def _remap_keypoints(kp, scale, pad_top, pad_left, offset_x=0, offset_y=0):
+            """Remap keypoints from model space back to original image space."""
+            kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
+            invalid = kp[..., 0] < 0
+            kp[..., 0] = (kp[..., 0] - pad_left) / scale + offset_x
+            kp[..., 1] = (kp[..., 1] - pad_top)  / scale + offset_y
+            kp[invalid] = -1
+            return kp
+
         def _run_on_latent(latent_batch):
             """Run one forward pass and return (keypoints_list, scores_list) for the batch."""
             nonlocal captured_feat
@@ -504,36 +528,19 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                         if x2 <= x1 or y2 <= y1:
                             continue
 
-                        crop_h_px, crop_w_px = y2 - y1, x2 - x1
                         crop = img[:, y1:y2, x1:x2, :]  # (1, crop_h, crop_w, C)
-
-                        # scale to fit inside (model_h, model_w) while preserving aspect ratio, then pad to exact model size.
-                        scale = min(model_h / crop_h_px, model_w / crop_w_px)
-                        scaled_h, scaled_w = int(round(crop_h_px * scale)), int(round(crop_w_px * scale))
-                        pad_top, pad_left  = (model_h - scaled_h) // 2, (model_w - scaled_w) // 2
-
-                        crop_chw = crop.permute(0, 3, 1, 2).float()  # BHWC → BCHW
-                        scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="bilinear", crop="disabled")
-                        padded = torch.zeros(1, scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
-                        padded[:, :, pad_top:pad_top + scaled_h, pad_left:pad_left + scaled_w] = scaled
-                        crop_resized = padded.permute(0, 2, 3, 1)  # BCHW → BHWC
+                        crop_resized, scale, pad_top, pad_left = _resize_to_model(crop)
 
                         latent_crop = vae.encode(crop_resized)
                         kp_batch, sc_batch = _run_on_latent(latent_crop)
-                        kp, sc = kp_batch[0], sc_batch[0]  # (K, 2), coords in model pixel space
-
-                        # remove padding offset, undo scale, offset to full-image coordinates.
-                        kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
-                        kp[..., 0] = (kp[..., 0] - pad_left) / scale + x1
-                        kp[..., 1] = (kp[..., 1] - pad_top)  / scale + y1
-
+                        kp = _remap_keypoints(kp_batch[0], scale, pad_top, pad_left, x1, y1)
                         img_keypoints.append(kp)
-                        img_scores.append(sc)
+                        img_scores.append(sc_batch[0])
                 else:
-                    # No bboxes for this image – run on the full image
-                    latent_img = vae.encode(img)
+                    img_resized, scale, pad_top, pad_left = _resize_to_model(img)
+                    latent_img = vae.encode(img_resized)
                     kp_batch, sc_batch = _run_on_latent(latent_img)
-                    img_keypoints.append(kp_batch[0])
+                    img_keypoints.append(_remap_keypoints(kp_batch[0], scale, pad_top, pad_left))
                     img_scores.append(sc_batch[0])
 
                 all_keypoints.append(img_keypoints)
@@ -541,19 +548,16 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                 pbar.update(1)
 
         else: # full-image mode, batched
-            tqdm_pbar = tqdm(total=total_images, desc="Extracting keypoints")
-            for batch_start in range(0, total_images, batch_size):
-                batch_end = min(batch_start + batch_size, total_images)
-                latent_batch = vae.encode(image[batch_start:batch_end])
-
+            for batch_start in tqdm(range(0, total_images, batch_size), desc="Extracting keypoints"):
+                batch_resized, scale, pad_top, pad_left = _resize_to_model(image[batch_start:batch_start + batch_size])
+                latent_batch = vae.encode(batch_resized)
                 kp_batch, sc_batch = _run_on_latent(latent_batch)
 
                 for kp, sc in zip(kp_batch, sc_batch):
-                    all_keypoints.append([kp])
+                    all_keypoints.append([_remap_keypoints(kp, scale, pad_top, pad_left)])
                     all_scores.append([sc])
-                    tqdm_pbar.update(1)
 
-                pbar.update(batch_end - batch_start)
+                pbar.update(len(kp_batch))
 
         openpose_frames = _to_openpose_frames(all_keypoints, all_scores, height, width)
         return io.NodeOutput(openpose_frames)

comfy_extras/nodes_upscale_model.py

@@ -6,6 +6,7 @@ import comfy.utils
 import folder_paths
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
+import comfy.model_management
 
 try:
     from spandrel_extra_arches import EXTRA_REGISTRY
@@ -78,13 +79,15 @@ class ImageUpscaleWithModel(io.ComfyNode):
         tile = 512
         overlap = 32
 
+        output_device = comfy.model_management.intermediate_device()
+
         oom = True
         try:
             while oom:
                 try:
                     steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap)
                     pbar = comfy.utils.ProgressBar(steps)
-                    s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar)
+                    s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a.float()), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar, output_device=output_device)
                     oom = False
                 except Exception as e:
                     model_management.raise_non_oom(e)
@@ -94,7 +97,7 @@ class ImageUpscaleWithModel(io.ComfyNode):
         finally:
             upscale_model.to("cpu")
 
-        s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0)
+        s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0).to(comfy.model_management.intermediate_dtype())
         return io.NodeOutput(s)
 
     upscale = execute  # TODO: remove

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.42.8
-comfyui-workflow-templates==0.9.44
+comfyui-frontend-package==1.42.10
+comfyui-workflow-templates==0.9.45
 comfyui-embedded-docs==0.4.3
 torch
 torchsde
@@ -33,5 +33,5 @@ kornia>=0.7.1
 spandrel
 pydantic~=2.0
 pydantic-settings~=2.0
-PyOpenGL>=3.1.8
-comfy-angle
+PyOpenGL
+glfw