feat: add MiniMax-H3-Image, fix reference image slot indices, bump requirements

README.md

@@ -57,6 +57,7 @@ models:
   - Comfy-Org/Lumina_Image_2.0_Repackaged
   - Comfy-Org/Mage-Flow
   - Comfy-Org/Ming-Image
+  - Comfy-Org/MiniMax-H3-Image
   - Comfy-Org/MOSS-Image
   - Comfy-Org/NewBie-image-Exp0.1_repackaged
   - Comfy-Org/Omnigen2_ComfyUI_repackaged
@@ -187,6 +188,7 @@ models:
   - meituan-longcat/LongCat-Image
   - microsoft/Lens
   - microsoft/Lens-Turbo
+  - MiniMaxAI/MiniMax-H3-Image
   - NewBie-AI/NewBie-image-Exp0.1
   - nvidia/PiD
   - nvidia/PixelDiT-1300M-1024px

chain_injectors/reference_image_injector.py

@@ -12,6 +12,80 @@ def create_node(assembler, class_type, title):
     node['_meta']['title'] = title
     return node
 
+
+def _get_node_input_spec(class_type, input_name):
+    """Fetch the raw input spec (``[type, config]``) for ``class_type.input_name``."""
+    # 1. ComfyUI /object_info backed manager (used by the standalone comfy-webui app)
+    try:
+        from comfy_webui_core import node_info_manager
+        node_info = node_info_manager.get_node_info(class_type)
+        if isinstance(node_info, dict):
+            inputs_cfg = node_info.get('input', {}) or {}
+            for category in ('required', 'optional'):
+                candidate = (inputs_cfg.get(category) or {}).get(input_name)
+                if candidate:
+                    return candidate
+    except Exception:
+        pass
+
+    # 2. In-process ComfyUI node classes (used by the Hugging Face Spaces runtime)
+    try:
+        from comfy_integration.nodes import NODE_CLASS_MAPPINGS
+        node_class = NODE_CLASS_MAPPINGS.get(class_type)
+        if node_class is not None:
+            input_types = node_class.INPUT_TYPES() or {}
+            for category in ('required', 'optional'):
+                candidate = (input_types.get(category) or {}).get(input_name)
+                if candidate:
+                    return candidate
+    except Exception:
+        pass
+
+    return None
+
+
+def _get_autogrow_image_slot_names(assembler, target_node_id, input_name='images'):
+    """
+    Resolve the real per-image slot key template for an Autogrow image input.
+
+    ComfyUI exposes two flavours of Autogrow template:
+      * TemplatePrefix(prefix="image_") -> names are 0-based: image_0, image_1, ...
+      * TemplateNames(names=["image_1", ...]) -> names start at image_1
+
+    Returns a list of slot names indexed from 0 (i.e. index i is the name for the i-th image),
+    or None when the node schema cannot be introspected.
+    """
+    target_node = assembler.workflow.get(target_node_id)
+    if not isinstance(target_node, dict):
+        return None
+
+    class_type = target_node.get('class_type')
+    if not class_type:
+        return None
+
+    spec = _get_node_input_spec(class_type, input_name)
+    if not isinstance(spec, (list, tuple)) or len(spec) < 2 or not isinstance(spec[1], dict):
+        return None
+
+    template = spec[1].get('template')
+    if not isinstance(template, dict):
+        return None
+
+    names = template.get('names')
+    if isinstance(names, (list, tuple)) and names:
+        return [str(n) for n in names]
+
+    prefix = template.get('prefix')
+    if isinstance(prefix, str):
+        # TemplatePrefix is numbered from 0 in ComfyUI (range(self.max)).
+        max_count = template.get('max', 100)
+        if not isinstance(max_count, int) or max_count <= 0:
+            max_count = 100
+        return [f"{prefix}{i}" for i in range(max_count)]
+
+    return None
+
+
 def inject(assembler, chain_definition, chain_items):
     if not chain_items:
         return
@@ -43,7 +117,7 @@ def inject(assembler, chain_definition, chain_items):
     target_node_id = assembler.node_map.get(target_node_name)
 
     if not target_node_id or target_node_id not in assembler.workflow:
-        supported_types = ('TextEncodeMageFlowEdit', 'TextEncodeQwenImage21', 'TextEncodeBooguEdit', 'TextEncodeMingImageEdit', 'TextEncodeFibo2')
+        supported_types = ('TextEncodeMageFlowEdit', 'TextEncodeQwenImage21', 'TextEncodeBooguEdit', 'TextEncodeMingImageEdit', 'TextEncodeFibo2', 'MiniMaxH3ImageConditioning')
         for node_id, node in assembler.workflow.items():
             if isinstance(node, dict) and node.get('class_type') in supported_types:
                 target_node_id = node_id
@@ -69,7 +143,24 @@ def inject(assembler, chain_definition, chain_items):
     megapixels = chain_definition.get('megapixels', 1.0)
     resolution_steps = chain_definition.get('resolution_steps', 1)
 
+    # Derive the node's real autogrow slot naming (0-based prefix vs 1-based names)
+    # instead of assuming image_1..image_N, which silently breaks 0-based nodes.
+    input_name = chain_definition.get('input_name', 'images')
+    slot_names = _get_autogrow_image_slot_names(assembler, target_node_id, input_name=input_name)
+    index_base = chain_definition.get('slot_index_base')
+    slot_key = None
+
     for i, img_filename in enumerate(valid_images):
+        if slot_names is not None:
+            if i >= len(slot_names):
+                print(f"Warning: Reference Image chain reached the node's slot limit ({len(slot_names)}). Stopping.")
+                break
+            slot_key = f"{input_name}.{slot_names[i]}"
+        else:
+            # Fallback when node schema is unavailable: honor explicit chain config or use 1-based legacy default.
+            base = 1 if index_base is None else int(index_base)
+            slot_key = f"{input_name}.image_{i + base}"
+
         load_id = assembler._get_unique_id()
         load_node = create_node(assembler, "LoadImage", f"Load Reference Image {i+1}")
         load_node['inputs']['image'] = img_filename
@@ -83,8 +174,7 @@ def inject(assembler, chain_definition, chain_items):
         scale_node['inputs']['image'] = [load_id, 0]
         assembler.workflow[scale_id] = scale_node
 
-        input_key = f"images.image_{i+1}"
-        assembler.workflow[target_node_id]['inputs'][input_key] = [scale_id, 0]
+        assembler.workflow[target_node_id]['inputs'][slot_key] = [scale_id, 0]
 
     node_class = assembler.workflow[target_node_id].get('class_type', 'Unknown')
-    print(f"Reference Image injector applied: {len(valid_images)} image(s) injected into '{target_node_id}' ({node_class}), max_images={max_images}.")
+    print(f"Reference Image injector applied: {len(valid_images)} image(s) injected into '{target_node_id}' ({node_class}), max_images={max_images}, slots={slot_key or '-'}.")

core/pipelines/workflow_recipes/_partials/conditioning/minimax-h3-image.yaml

@@ -0,0 +1,72 @@
+nodes:
+  unet_loader:
+    class_type: UNETLoader
+    title: "Load Diffusion Model"
+    params:
+      weight_dtype: "default"
+  vae_loader:
+    class_type: VAELoader
+    title: "Load VAE"
+  clip_loader:
+    class_type: CLIPLoader
+    title: "Load CLIP"
+    params:
+      type: "minimax"
+      device: "default"
+  prompt:
+    class_type: MiniMaxH3ImageConditioning
+    title: "MiniMax H3 Image Conditioning"
+    params:
+      width: 2048
+      height: 2048
+  negative_zero_out:
+    class_type: ConditioningZeroOut
+    title: "Conditioning Zero Out (Negative)"
+
+connections:
+  - from: "unet_loader:0"
+    to: "ksampler:model"
+  - from: "clip_loader:0"
+    to: "prompt:clip"
+  - from: "vae_loader:0"
+    to: "prompt:vae"
+  - from: "prompt:0"
+    to: "ksampler:positive"
+  - from: "prompt:0"
+    to: "negative_zero_out:conditioning"
+  - from: "negative_zero_out:0"
+    to: "ksampler:negative"
+  - from: "prompt:1"
+    to: "ksampler:latent_image"
+  - from: "vae_loader:0"
+    to: "vae_decode:vae"
+  - from: "vae_loader:0"
+    to: "vae_encode:vae"
+
+dynamic_lora_chains:
+  lora_chain:
+    template: "LoraLoader"
+    output_map:
+      "unet_loader:0": "model"
+      "clip_loader:0": "clip"
+    input_map:
+      "model": "model"
+      "clip": "clip"
+    end_input_map:
+      "model": ["ksampler:model"]
+      "clip": ["prompt:clip"]
+
+dynamic_reference_image_chains:
+  reference_image_chain:
+    text_encode_node: "prompt"
+    vae_node: "vae_loader"
+    max_images: 8
+
+ui_map:
+  positive_prompt: "prompt:prompt"
+  negative_prompt: "dummy:negative_prompt"
+  unet_name: "unet_loader:unet_name"
+  clip_name: "clip_loader:clip_name"
+  vae_name: "vae_loader:vae_name"
+  width: "prompt:width"
+  height: "prompt:height"

requirements.txt

@@ -1,6 +1,6 @@
-comfyui-frontend-package==1.57.1
-comfyui-workflow-templates==0.11.96
-comfyui-embedded-docs==0.5.14
+comfyui-frontend-package==1.58.5
+comfyui-workflow-templates==0.11.106
+comfyui-embedded-docs==0.5.16
 torch
 torchsde
 torchvision
@@ -22,7 +22,7 @@ SQLAlchemy>=2.0.0
 filelock
 av>=17.0.0
 comfy-aimdo==0.6.0
-comfy-kitchen==0.2.40
+comfy-kitchen==0.2.41
 requests
 simpleeval>=1.0.0
 blake3

yaml/chain_features.yaml

@@ -251,6 +251,21 @@ fibo2_reference_edit:
     required:
       - image
 
+minimax_h3_image_reference_edit:
+  chains: reference_image
+  display_name: "MiniMax-H3-Image Reference Edit"
+  description: "For MiniMax-H3-Image, this feature enables reference image editing and combining capabilities. In txt2img mode, adding a single reference image performs an Image Edit, while adding multiple images performs an Image Combine. The prompt refers to them as <Picture 1>, <Picture 2>, ..."
+  max_count: 8
+  usage_guideline: "Supply reference image(s) encoded as Base64 Data URI or HTTP/HTTPS URL (up to 8). Refer to them in the prompt as <Picture 1>, <Picture 2>, ...; <Picture 1> is the picture to edit."
+  parameters_schema:
+    type: object
+    properties:
+      image:
+        type: string
+        description: "Reference image encoded as Base64 Data URI or HTTP/HTTPS URL."
+    required:
+      - image
+
 krea2_identity_edit:
   chains: krea2_identity_edit
   display_name: "KREA2 Identity Edit"

yaml/constants.yaml

@@ -24,6 +24,14 @@ RESOLUTION_MAP:
     "3:4 (Classic Portrait)": [896, 1152]
     "3:2 (Photography)": [1216, 832]
     "2:3 (Photography Portrait)": [832, 1216]
+  minimax-h3-image:
+    "1:1 (Square)": [2048, 2048]
+    "16:9 (Landscape)": [2048, 1152]
+    "9:16 (Portrait)": [1152, 2048]
+    "4:3 (Classic)": [1792, 1344]
+    "3:4 (Classic Portrait)": [1344, 1792]
+    "3:2 (Photography)": [1920, 1280]
+    "2:3 (Photography Portrait)": [1280, 1920]
   krea-2:
     "1:1 (Square)": [1024, 1024]
     "16:9 (Landscape)": [1344, 768]

yaml/file_list.yaml

@@ -405,6 +405,11 @@ file:
       repo_id: "comfyanonymous/ControlNet-v1-1_fp16_safetensors"
       repository_file_path: "control_v11u_sd15_tile_fp16.safetensors"
   diffusion_models: 
+    # MiniMax-H3-Image
+    - filename: "minimax_h3_image_pruned_w4a8.safetensors"
+      source: "hf"
+      repo_id: "Comfy-Org/MiniMax-H3-Image"
+      repository_file_path: "diffusion_models/minimax_h3_image_pruned_w4a8.safetensors"
     # Fibo 2.0
     - filename: "fibo2_dev_int8_convrot.safetensors"
       source: "hf"
@@ -970,6 +975,11 @@ file:
       repo_id: "Comfy-Org/Qwen-Image-2.1"
       repository_file_path: "model_patches/qwen_image_2.1_fun_controlnet_union_int8_convrot.safetensors"
   text_encoders: 
+    # MiniMax-H3-Image
+    - filename: "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
+      source: "hf"
+      repo_id: "Comfy-Org/MiniMax-H3-Image"
+      repository_file_path: "text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
     # MOSS-Image
     - filename: "qwen3.5_9b_w4a8_gptq.safetensors"
       source: "hf"
@@ -1119,6 +1129,11 @@ file:
       repo_id: "Comfy-Org/Omnigen2_ComfyUI_repackaged"
       repository_file_path: "split_files/text_encoders/qwen_2.5_vl_fp16.safetensors"  
   vae: 
+    # MiniMax-H3-Image
+    - filename: "minimax_h3_image_vae_fp16.safetensors"
+      source: "hf"
+      repo_id: "Comfy-Org/MiniMax-H3-Image"
+      repository_file_path: "vae/minimax_h3_image_vae_fp16.safetensors"
     - filename: "wan_2.1_vae.safetensors"
       source: "hf"
       repo_id: "Comfy-Org/Wan_2.1_ComfyUI_repackaged"

yaml/model_architecture_features.yaml

@@ -12,6 +12,11 @@ flux3:
   - flux3_prompt_upsampler
   - pid
 
+minimax-h3-image:
+  enabled_chains:
+  - lora
+  - reference_image
+
 krea-2:
   enabled_chains:
   - lora

yaml/model_architectures.yaml

@@ -1,5 +1,6 @@
 architecture_order:
   - "FLUX.3"
+  - "MiniMax-H3-Image"
   - "Krea-2"
   - "Qwen-Image-2.1"
   - "Fibo-2.0"
@@ -43,6 +44,8 @@ architectures:
     model_type: "fibo-2.0"
   "FLUX.3":
     model_type: "flux3"
+  "MiniMax-H3-Image":
+    model_type: "minimax-h3-image"
   Krea-2:
     model_type: "krea-2"
   "SenseNova-U1.5":

yaml/model_defaults.yaml

@@ -45,6 +45,13 @@ FLUX.3:
     sampler_name: "euler"
     scheduler: "simple"
 
+MiniMax-H3-Image:
+  _defaults:
+    steps: 6
+    cfg: 1.0
+    sampler_name: "euler"
+    scheduler: "simple"
+
 Krea-2:
   _defaults:
     steps: 52

yaml/model_list.yaml

@@ -17,6 +17,14 @@ Checkpoint:
           unet: "flux3-image-klein-base-nvfp4.safetensors"
           clip: "qwen3vl_4b_fp8_scaled.safetensors"
           vae: "flux2-vae.safetensors"
+  MiniMax-H3-Image:
+    latent_type: none
+    models:
+      - display_name: "MiniMaxAI/MiniMax-H3-Image"
+        components:
+          unet: "minimax_h3_image_pruned_w4a8.safetensors"
+          clip: "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
+          vae: "minimax_h3_image_vae_fp16.safetensors"
   Krea-2:
     latent_type: latent
     models: