# Copyright (c) 2026 DeepBeepMeep - part of WanGP, WanGP Community License 2.0 (see LICENSE.txt) """WanGP integration for inclusionAI Ming-Image-0.1-Design. The checkpoint URLs describe the prepared DeepBeepMeep layout. The upstream code and checkpoint provenance are recorded in README.md. """ from pathlib import Path import torch from shared.utils import files_locator as fl from shared.utils.hf import build_hf_url from .prompt_enhancers import CLASSIC_IMAGE, CLASSIC_TEXT, JSON_IMAGE, JSON_TEXT, LAYER_PLAN REPO = "DeepBeepMeep/MingImage" PROJECT = "ming_image" ENCODER = "BailingMM2-Ming-Image" ENCODER_CORE = "BailingMM2-Ming-Image-Core" SHARED_PROJECT = "ming_image_shared" ARCHITECTURE = "ming_image_0_1_design" TRANSFORMER = "Ming-Image-0.1-Design" LAYER_REPO = REPO LAYER_PROJECT = "ming_image_layer" LAYER_ARCHITECTURE = "ming_image_0_1_design_layer" LAYER_TRANSFORMER = "Ming-Image-0.1-Design-Layer" TOKENIZER_FILES = ( "config.json", "preprocessor_config.json", "special_tokens_map.json", "tokenizer.json", "tokenizer_config.json", ) STRUCTURED_PROMPT_EXAMPLE = """{ "canvas_settings": { "aspect_ratio": "1:1, 2048 × 2048 px", "ambient_lighting": "Soft, even studio light", "image_style": "Clean modern editorial poster" }, "layers": [ { "description": "Full-canvas deep blue background with subtle paper texture.", "coordinates": "cx: 0.500, cy: 0.500, w: 1.000, h: 1.000", "hierarchy_and_relation": "Backmost layer covering the canvas.", "color_specs": ["#14213D"] }, { "description": "Large centered white title \\"OPEN HOUSE\\" in bold sans-serif lettering.", "coordinates": "cx: 0.500, cy: 0.250, w: 0.700, h: 0.160", "hierarchy_and_relation": "Top layer centered above the main artwork.", "color_specs": ["#FFFFFF"] } ] }""" class family_handler: @staticmethod def resolve_runtime_model_def(model_def, runtime_context): variant = "int8_convrot" if runtime_context.get("text_encoder_quantization") == "int8" else "bf16" return {**model_def, "ming_encoder_variant": variant} @staticmethod def render_prompt_helper(model_type, model_def, prompt_id, popup_id, prompt_elem_id, resolution_elem_id): if model_type == LAYER_ARCHITECTURE: return "" from .prompt_helper import render_prompt_helper return render_prompt_helper(model_type, model_def, prompt_id, popup_id, prompt_elem_id, resolution_elem_id) @staticmethod def get_prompt_helper_css(): from .prompt_helper import get_prompt_helper_css return get_prompt_helper_css() @staticmethod def get_prompt_helper_javascript(): from .prompt_helper import get_prompt_helper_javascript return get_prompt_helper_javascript() @staticmethod def query_supported_types(): return [ARCHITECTURE, LAYER_ARCHITECTURE] @staticmethod def query_model_family(): return "ming_image" @staticmethod def query_family_infos(): return {"ming_image": (1160, "Ming Image 0.1")} @staticmethod def query_family_maps(): return {}, {} @staticmethod def get_lora_dir(base_model_type): return LAYER_PROJECT if base_model_type == LAYER_ARCHITECTURE else PROJECT @staticmethod def get_rgb_factors(base_model_type): from .rgb_factors import RGB_BIAS, RGB_FACTORS return RGB_FACTORS, RGB_BIAS @staticmethod def query_model_def(base_model_type, model_def): if base_model_type == LAYER_ARCHITECTURE: return { "image_outputs": True, "device_explicit": True, "skip_prompt_template": True, "embedded_guidance": False, "guidance_max_phases": 1, "NAG": False, "profiles_dir": [LAYER_ARCHITECTURE], "resolution_budget": "1024x1024", "text_encoder_folder": ENCODER, "text_encoder_URLs": [ build_hf_url(REPO, ENCODER, ENCODER_CORE + suffix) for suffix in ("_bf16.safetensors", "_int8_convrot.safetensors") ], "image_ref_choices": { "choices": [("Decompose the Reference Image", "I")], "letters_filter": "I", "default": "I", }, "one_image_ref_needed": True, "one_image_ref_only": True, "no_background_removal": True, "prompt_enhancer_def": { "selection": ["TI"], "labels": {"TI": "Build a Layer Plan from Text Prompt and {image_inputs}"}, "default": "", }, "image_prompt_enhancer_instructions": LAYER_PLAN, "image_prompt_enhancer_max_tokens": 1024, "specialities": [{ "name": "layer decomposition", "aliases": ["transparent design layers"], "description": "Separates one flattened design into ordered transparent PNG layers.", }, {"name": "alpha output", "aliases": ["RGBA", "transparent layers"], "description": "Separate a supplied design into transparent PNG layers."}], "infos": ( "**Ming Image Design-Layer** turns a finished poster, infographic, card, or similar design " "into separate transparent image layers. Select one Reference Image, then describe the pieces " "you want to separate in Text Prompt, starting with the frontmost. For example, put the words " "on one layer, the panel behind them on another, the illustration on a third, and the " "background last. Describe only elements that are visible in the image.\n\n" "WanGP adds every layer to the gallery and saves them together in a ZIP. You can hide or " "rearrange the layers in an image editor; text remains part of the image rather than editable " "type. Start with 12 steps, guidance 2, and 1024x1024 for a detailed square result. " "Choose 512x512 when speed matters more than detail." ), "prompt_infos": ( "Write `Decompose this image into N layers with the following specifications:` followed by " "`Number of layers: N` and one `Layer i:` line for each layer. Order layers front to back; " "put readable text on a front layer, any card or panel directly behind it on its own layer, " "the main subject on its own layer, and the background last. Quote actual visible words " "exactly and do not invent unreadable lettering. For a quick request, `Decompose this image " "into 4 layers.` also works.\n\n" "**Example:**\n```text\nDecompose this image into 4 layers with the following specifications:\n" "Number of layers: 4\nLayer 1: Exact visible headline and subtitle on transparent pixels.\n" "Layer 2: The colored title card directly behind the text.\n" "Layer 3: The main foreground illustration.\n" "Layer 4: The complete background and remaining supporting shapes.\n```" ), "deepy_infos": ( "Ming Image Design-Layer separates one supplied design image into transparent layers. " "Ask for the visible pieces from front to back, with the background last. Each layer appears " "in the gallery and in a ZIP. Lettering remains image pixels. Start with 12 steps, guidance 2, " "and 1024x1024; choose 512x512 for a faster draft." ), "deepy_prompt_infos": ( "Ask for `Decompose this image into N layers` or supply `Number of layers: N` and exact " "`Layer 1:` through `Layer N:` descriptions. The first layer is topmost and the last is " "the background. Ground every described object and quoted string in the reference image. " "Place text, its supporting panel, the main subject, and the background on separate layers " "when present." ), } return { "image_outputs": True, "device_explicit": True, "skip_prompt_template": True, "prompt_helper_popup_dims": [86, 94], "prompt_enhancer_def": { "selection": ["T", "TI", "T1", "TI1"], "labels": { "T": "A Classic Image Prompt using existing Text Prompt", "TI": "A Classic Image Edit Prompt using existing Text Prompt and {image_inputs}", "T1": "A Ming JSON Image Prompt using existing Text Prompt", "TI1": "A Ming JSON Image Edit Prompt using existing Text Prompt and {image_inputs}", }, "default": "", }, "text_prompt_enhancer_instructions": CLASSIC_TEXT, "image_prompt_enhancer_instructions": CLASSIC_IMAGE, "text_prompt_enhancer_instructions1": JSON_TEXT, "image_prompt_enhancer_instructions1": JSON_IMAGE, "text_prompt_enhancer_max_tokens": 1024, "image_prompt_enhancer_max_tokens": 1024, "text_prompt_enhancer_max_tokens1": 4096, "image_prompt_enhancer_max_tokens1": 4096, "embedded_guidance": False, "NAG": False, "profiles_dir": [ARCHITECTURE], "resolution_budget": "2048x2048", "text_encoder_folder": ENCODER, "text_encoder_URLs": [ build_hf_url(REPO, ENCODER, ENCODER_CORE + suffix) for suffix in ("_bf16.safetensors", "_int8_convrot.safetensors") ], "image_ref_choices": { "choices": [ ("None", ""), ("Edit the Reference Image", "I"), ], "letters_filter": "I", }, "one_image_ref_only": True, "no_background_removal": True, "custom_settings": [{ "id": "rgba", "name": "RGBA", "label": "RGBA Output", "type": "dropdown", "default": "Disabled", "choices": [("Disabled", "Disabled"), ("Enabled", "Enabled")], "info": "Preserve the model-generated alpha channel in PNG. Transparency depends on the prompt and is not guaranteed.", }], "specialities": [ {"name": "alpha output", "aliases": ["RGBA", "transparent background"], "description": "Set custom_settings.rgba to Enabled and request a transparent background to preserve model-generated alpha in PNG. Transparency depends on the prompt."}, {"name": "visual design", "aliases": ["UI design", "poster design"], "description": "Creates text-rich visual designs, posters, and interface mockups."}, ], "infos": ( "**Ming Image Design** creates posters, infographics, interface mockups, and other visual designs " "from text. You can also select one Reference Image and describe a change while saying what to " "keep. For a simple image, write a normal prompt. For a precise layout with titles and labels, " "use the Ming JSON prompt enhancer or arrange the design in Prompt Helper. Put any words that " "must appear exactly in quotes.\n\n" "Start with 12 steps and guidance 1. The default 2048x2048 size gives more detail; choose " "1024x1024 to save time and memory. Reference edits work best at 1024x1024; larger edit " "outputs are resized. To request transparency, enable RGBA Output and ask for a transparent " "background, though the result may still be opaque. For separating an existing design into " "individual layers, choose Ming Image Design-Layer instead." ), "prompt_infos": ( "For text-to-image designs with exact layout, paste one JSON object into Prompt. It has exactly two top-level " "keys: `canvas_settings` (`aspect_ratio`, `ambient_lighting`, `image_style`) and `layers` (backmost to " "frontmost). Each layer has `description`, `coordinates`, `hierarchy_and_relation`, and `color_specs` " "(an array of hex colors). Write `coordinates` as one normalized string: " "`cx: 0.500, cy: 0.500, w: 1.000, h: 1.000`. Keep each box inside the canvas, describe one complete " "visible group per layer, and quote every exact rendered string once in its owning layer's description. " "Match the JSON aspect ratio to the selected WanGP resolution. Use Prompt Helper to draw or edit layer boxes; " "it updates the size field from the selected resolution.\n\n" "**Example:**\n```json\n" + STRUCTURED_PROMPT_EXAMPLE + "\n```\n\n" "Plain-language prompts also work. For image editing, describe the change and what to preserve, for " "example: Change the headline to \"OPEN STUDIO\" while keeping the layout and colors. For transparency, " "enable RGBA Output and begin a plain-language prompt with `RGBA, 4-channel, transparent background`." ), "deepy_infos": ( "Ming Image Design creates posters, infographics, and other visual designs from text or edits " "one reference image. Use a plain prompt for simple work or structured JSON for precise " "layouts. Quote exact visible words. For an edit, name the change and what " "to preserve. Start with 12 steps and guidance 1; 2048x2048 gives more detail, while " "1024x1024 is faster and suits edits. For transparency, enable RGBA Output and request a " "transparent background. Use Design-Layer to separate an existing image into layers." ), "deepy_prompt_infos": ( "For a text-to-image design needing exact layout or lettering, write one JSON object with exactly " "`canvas_settings` (`aspect_ratio`, `ambient_lighting`, `image_style`) and `layers` in back-to-front order. " "Each visible semantic layer contains exactly `description`, `coordinates`, `hierarchy_and_relation`, " "and `color_specs` (hex-color array). `coordinates` is one normalized string in the form " "`cx: 0.500, cy: 0.500, w: 1.000, h: 1.000`; boxes stay inside the canvas. Keep people and objects " "intact, use the fewest groups that preserve the layout, and put each exact visible string once, quoted " "verbatim, in its owning description. Use relationships for alignment, containment, stacking and " "occlusion, with no duplicate text or invisible layers. Match `aspect_ratio` to the chosen WanGP output " "size. For simple requests, a concrete plain-language prompt is valid. For reference editing, use a " "specific plain-language change plus what must remain unchanged. Do not request layer decomposition." ), } @staticmethod def query_model_files(computeList, base_model_type, model_def=None): variant = (model_def or {}).get("ming_encoder_variant", "bf16") vision_file = f"vision_encoder_{variant}.safetensors" conditioner_file = f"conditioning_{variant}.safetensors" if base_model_type == LAYER_ARCHITECTURE: return [ {"repoId": LAYER_REPO, "sourceFolderList": [LAYER_PROJECT], "fileList": [[ "transformer_config.json", "connector_config.json", "mlp_config.json", "scheduler_config.json", conditioner_file, ]]}, {"repoId": REPO, "sourceFolderList": [SHARED_PROJECT], "fileList": [[vision_file]]}, {"repoId": REPO, "sourceFolderList": [ENCODER], "fileList": [list(TOKENIZER_FILES)]}, {"repoId": REPO, "sourceFolderList": [PROJECT], "fileList": [[ "vae_config.json", "vae.safetensors", ]]}, ] return [ {"repoId": REPO, "sourceFolderList": [PROJECT], "fileList": [[ "transformer_config.json", "connector_config.json", "mlp_config.json", "vae_config.json", "vae.safetensors", "scheduler_config.json", conditioner_file, ]]}, {"repoId": REPO, "sourceFolderList": [SHARED_PROJECT], "fileList": [[vision_file]]}, {"repoId": REPO, "sourceFolderList": [ENCODER], "fileList": [list(TOKENIZER_FILES)]}, ] @staticmethod def update_default_settings(base_model_type, model_def, ui_defaults): if base_model_type == LAYER_ARCHITECTURE: ui_defaults.update(image_mode=1, resolution="1024x1024", num_inference_steps=12, guidance_scale=2.0, batch_size=1, prompt_enhancer="", video_prompt_type="I", remove_background_images_ref=0) return ui_defaults.update(image_mode=1, resolution="2048x2048", num_inference_steps=12, guidance_scale=1.0, batch_size=1, prompt_enhancer="", video_prompt_type="", remove_background_images_ref=0) @staticmethod def load_model(model_filename, model_type, base_model_type, model_def, text_encoder_filename=None, save_quantized=False, quantizeTransformer=False, **kwargs): from .ming_pipeline import load_components return load_components( transformer_filename=model_filename[0], text_encoder_filename=text_encoder_filename, model_type=model_type, save_quantized=save_quantized, quantize_transformer=quantizeTransformer, model_def=model_def, )