From 027c63f63a7f5f380a4df1057c548410b0a87606 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 15 Aug 2025 21:57:47 +0300 Subject: [PATCH 1/3] fix(OpenAIGPTImage1): set correct MIME type for multipart uploads to OpenAI edits (#9348) --- comfy_api_nodes/nodes_openai.py | 13 +++---------- 1 file changed, 3 insertions(+), 10 deletions(-) diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index ab3c5363bad7..cbff2b2d2f61 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -464,8 +464,6 @@ async def api_call( path = "/proxy/openai/images/generations" content_type = "application/json" request_class = OpenAIImageGenerationRequest - img_binaries = [] - mask_binary = None files = [] if image is not None: @@ -484,14 +482,11 @@ async def api_call( img_byte_arr = io.BytesIO() img.save(img_byte_arr, format="PNG") img_byte_arr.seek(0) - img_binary = img_byte_arr - img_binary.name = f"image_{i}.png" - img_binaries.append(img_binary) if batch_size == 1: - files.append(("image", img_binary)) + files.append(("image", (f"image_{i}.png", img_byte_arr, "image/png"))) else: - files.append(("image[]", img_binary)) + files.append(("image[]", (f"image_{i}.png", img_byte_arr, "image/png"))) if mask is not None: if image is None: @@ -511,9 +506,7 @@ async def api_call( mask_img_byte_arr = io.BytesIO() mask_img.save(mask_img_byte_arr, format="PNG") mask_img_byte_arr.seek(0) - mask_binary = mask_img_byte_arr - mask_binary.name = "mask.png" - files.append(("mask", mask_binary)) + files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png"))) # Build the operation operation = SynchronousOperation( From c308a8840aebf06649364e8e175862250a2d8823 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 15 Aug 2025 12:50:39 -0700 Subject: [PATCH 2/3] Add FluxKontextMultiReferenceLatentMethod node. (#9356) This node is only useful if someone trains the kontext model to properly use multiple reference images via the index method. The default is the offset method which feeds the multiple images like if they were stitched together as one. This method works with the current flux kontext model. --- comfy/ldm/flux/model.py | 26 +++++++++++++++++--------- comfy/model_base.py | 4 ++++ comfy_extras/nodes_flux.py | 19 +++++++++++++++++++ 3 files changed, 40 insertions(+), 9 deletions(-) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index 8f4d99f548f3..c4de82795b17 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -224,19 +224,27 @@ def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, if ref_latents is not None: h = 0 w = 0 + index = 0 + index_ref_method = kwargs.get("ref_latents_method", "offset") == "index" for ref in ref_latents: - h_offset = 0 - w_offset = 0 - if ref.shape[-2] + h > ref.shape[-1] + w: - w_offset = w + if index_ref_method: + index += 1 + h_offset = 0 + w_offset = 0 else: - h_offset = h - - kontext, kontext_ids = self.process_img(ref, index=1, h_offset=h_offset, w_offset=w_offset) + index = 1 + h_offset = 0 + w_offset = 0 + if ref.shape[-2] + h > ref.shape[-1] + w: + w_offset = w + else: + h_offset = h + h = max(h, ref.shape[-2] + h_offset) + w = max(w, ref.shape[-1] + w_offset) + + kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset) img = torch.cat([img, kontext], dim=1) img_ids = torch.cat([img_ids, kontext_ids], dim=1) - h = max(h, ref.shape[-2] + h_offset) - w = max(w, ref.shape[-1] + w_offset) txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None)) diff --git a/comfy/model_base.py b/comfy/model_base.py index cde61df7c925..bf874b875c88 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -890,6 +890,10 @@ def extra_conds(self, **kwargs): for lat in ref_latents: latents.append(self.process_latent_in(lat)) out['ref_latents'] = comfy.conds.CONDList(latents) + + ref_latents_method = kwargs.get("reference_latents_method", None) + if ref_latents_method is not None: + out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method) return out def extra_conds_shapes(self, **kwargs): diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py index 8a8a1769801c..c8db75bb39df 100644 --- a/comfy_extras/nodes_flux.py +++ b/comfy_extras/nodes_flux.py @@ -100,9 +100,28 @@ def scale(self, image): return (image, ) +class FluxKontextMultiReferenceLatentMethod: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "conditioning": ("CONDITIONING", ), + "reference_latents_method": (("offset", "index"), ), + }} + + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "append" + EXPERIMENTAL = True + + CATEGORY = "advanced/conditioning/flux" + + def append(self, conditioning, reference_latents_method): + c = node_helpers.conditioning_set_values(conditioning, {"reference_latents_method": reference_latents_method}) + return (c, ) + NODE_CLASS_MAPPINGS = { "CLIPTextEncodeFlux": CLIPTextEncodeFlux, "FluxGuidance": FluxGuidance, "FluxDisableGuidance": FluxDisableGuidance, "FluxKontextImageScale": FluxKontextImageScale, + "FluxKontextMultiReferenceLatentMethod": FluxKontextMultiReferenceLatentMethod, } From 1702e6df16b0a52e147f19e3d5c5548c25a64339 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 15 Aug 2025 14:29:58 -0700 Subject: [PATCH 3/3] Implement wan2.2 camera model. (#9357) Use the old WanCameraImageToVideo node. --- comfy/ldm/wan/model.py | 7 ++++++- comfy/model_detection.py | 5 ++++- comfy/supported_models.py | 14 +++++++++++++- comfy_extras/nodes_wan.py | 7 +++++-- 4 files changed, 28 insertions(+), 5 deletions(-) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 4e2d995664f1..9d3741be3907 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -768,7 +768,12 @@ def __init__(self, operations=None, ): - super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations) + if model_type == 'camera': + model_type = 'i2v' + else: + model_type = 't2v' + + super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations) operation_settings = {"operations": operations, "device": device, "dtype": dtype} self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 8acc51e20e18..2bec0541ec6c 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -364,7 +364,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1] dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.') elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys: - dit_config["model_type"] = "camera" + if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys: + dit_config["model_type"] = "camera" + else: + dit_config["model_type"] = "camera_2.2" else: if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys: dit_config["model_type"] = "i2v" diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 156ff9e263bb..7ed6dfd69577 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1046,6 +1046,18 @@ class WAN21_Camera(WAN21_T2V): def get_model(self, state_dict, prefix="", device=None): out = model_base.WAN21_Camera(self, image_to_video=False, device=device) return out + +class WAN22_Camera(WAN21_T2V): + unet_config = { + "image_model": "wan2.1", + "model_type": "camera_2.2", + "in_dim": 36, + } + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.WAN21_Camera(self, image_to_video=False, device=device) + return out + class WAN21_Vace(WAN21_T2V): unet_config = { "image_model": "wan2.1", @@ -1260,6 +1272,6 @@ def clip_target(self, state_dict={}): return supported_models_base.ClipTarget(comfy.text_encoders.qwen_image.QwenImageTokenizer, comfy.text_encoders.qwen_image.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, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage] +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, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage] models += [SVD_img2vid] diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index 694a183f62c2..83a9906887df 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -422,9 +422,12 @@ def execute(cls, positive, negative, vae, width, height, length, batch_size, sta start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) concat_latent_image = vae.encode(start_image[:, :, :, :3]) concat_latent[:,:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]] + mask = torch.ones((1, 1, latent.shape[2] * 4, latent.shape[-2], latent.shape[-1])) + mask[:, :, :start_image.shape[0] + 3] = 0.0 + mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2) - positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent}) - negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent}) + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent, "concat_mask": mask}) if camera_conditions is not None: positive = node_helpers.conditioning_set_values(positive, {'camera_conditions': camera_conditions})