# Copyright (c) 2026 DeepBeepMeep - part of WanGP, WanGP Community License 2.0 (see LICENSE.txt) import torch from PIL import Image from shared.utils.hf import build_hf_url from .prompt_enhancers import SENSENOVA_GENERIC_PROMPT, SENSENOVA_GENERIC_REFERENCE_PROMPT, SENSENOVA_INFOGRAPHIC_PROMPT, SENSENOVA_INFOGRAPHIC_REFERENCE_PROMPT _ARCHITECTURE = "sensenova_u1_5_8b_mot" _PROJECT_REPO = "DeepBeepMeep/SenseNova" _PROJECT_FOLDER = "sensenova_u1_5" _PROFILE_FOLDER = _ARCHITECTURE SENSENOVA_INFOS = """## SenseNova-U1.5 SenseNova-U1.5 is a unified model for text-to-image generation and image editing. Generate from text with no reference image, or provide one or more ordered reference images for editing and reference-guided generation. ### Practical Settings - **Standard quality:** use `50` steps, **Guidance Scale** `4.0`, and **Flow Shift** `3.0`, matching the upstream reference settings. - **Fast generation:** load the **Official 8-Step Accelerator** profile, which selects the matching LoRA and step count. - **Resolution:** `2048x2048` is a good general-purpose choice. Native 4K improves space for detail and typography but requires substantially more time and memory. WanGP keeps both dimensions aligned to the model's 32-pixel image grid. - **Reference mode:** the default treats the first image as the main subject or landscape and derives output dimensions from it. Select **Use Reference Images** instead when every image is an ordinary reference and the chosen output resolution should be kept. - **Reference quality:** use clean, high-resolution source images when possible. WanGP sends multiple references in their displayed order, so describe each image's role in the prompt. ### Current Limitations - Dense, lengthy, very small, or mixed Chinese-English text can still contain spelling errors. - Exact counts, alignment, and hierarchy may drift in heavily constrained layouts. - Small faces, hands, limbs, and fine object structures can be unstable. - Broad or complex multi-reference edits may alter content that was meant to stay unchanged; state preservation requirements explicitly. - If details or colors become excessive, try lowering **Guidance Scale**. See the [official model card](https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT) and [SenseNova-U1.5 cookbook](https://github.com/OpenSenseNova/SenseNova-U1/blob/refs/heads/feat/u1.5/docs/u1.5_best_practices.md) for upstream examples and guidance.""" SENSENOVA_PROMPT_INFOS = """## SenseNova-U1.5 Prompt Guide ### Text-to-Image Direct natural language works well for a clear subject with few constraints. Describe the **subject**, **setting**, **composition**, **style**, **palette**, **materials**, **lighting**, and **camera or viewpoint** in the order that matters most. - State exact object counts and spatial relationships instead of relying on implication. - For visible text, put every literal string in double quotes and specify its language, placement, size, and typographic hierarchy. - For posters and infographics, organize the prompt into sections such as title, layout, panels, exact copy, icons, palette, typography, and constraints. - Use prompt enhancement for a short creative brief that needs a full design plan. Skip enhancement when the prompt is already long, structured, and production-ready. ### Prompt Enhancer Modes - **General Image Prompt:** expands a short brief into a balanced visual description without overloading a simple request. - **Infographic Prompt:** builds a visually led information design with semantic icons, illustrations, charts, diagrams, compact copy, and explicit typography. - Each mode also has a **First Reference Image** variant for edits or reference-guided subject, composition, layout, palette, and style reuse. ### Editing and References Put the requested change first, then state what must remain unchanged. - Name the edit target, location, quantity, color, material, size, or orientation precisely. - For text replacement, write both strings exactly: `Replace "OLD TEXT" with "NEW TEXT"`. - With multiple references, assign each a numbered role: `Use Image 1 as the base composition, Image 2 for the subject, and Image 3 for the color palette.` - For a local edit, explicitly preserve identity, pose, framing, lighting, background, and every untouched region that matters. - For a style reference, describe which visible traits to transfer—palette, texture, typography, lighting, spacing, or composition—rather than saying only `use this style`. ### Prompt Shapes - **Generation:** `A cinematic mountain lake at sunrise, realistic photography, calm reflections, natural mist, balanced wide composition.` - **Editing:** `Change the jacket to cobalt blue. Preserve the face, pose, background, lighting, and framing.` - **Infographic:** `Create a vertical infographic titled "CITY TREES". Use three clearly separated panels with the exact headings "COOLER STREETS", "CLEANER AIR", and "WILDLIFE", crisp sans-serif typography, restrained green palette, and no additional text.` Upstream recommends prompt enhancement especially for short infographic prompts because additional structure improves hierarchy, typography, and information density. See the [prompt-enhancement guide](https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/prompt_enhancement.md).""" class family_handler: @staticmethod def query_model_def(base_model_type, model_def): return { "image_outputs": True, "device_explicit": True, "cuda_graphs": ["transformer"], "dtype": "bf16", "no_negative_prompt": True, "no_background_removal": True, "guidance_max_phases": 1, "inference_steps": True, "flow_shift": True, "fit_into_canvas_image_refs": 0, "image_ref_choices": { "choices": [ ("Generate without Reference Images", ""), ("First Reference Image is the Main Subject / Landscape, defines Output Dimensions, and may be followed by other Reference Images", "KI"), ("Use Reference Images", "I"), ], "letters_filter": "KI", "default": "KI", "label": "Reference Images", }, "at_least_one_image_ref_needed": False, "image_prompt_types_allowed": "S", "infos": SENSENOVA_INFOS, "specialities": [{"name": "infographics", "aliases": ["information graphics", "information posters"]}, {"name": "text rendering", "description": "Chinese and English lettering within images."}, {"name": "precise layouts"}], "prompt_infos": SENSENOVA_PROMPT_INFOS, "deepy_infos": "### SenseNova-U1.5\n\nText-to-image and editing with ordered references. First-reference main-scene mode derives dimensions from that image; ordinary-reference mode keeps chosen resolution. Start at 50 steps, guidance 4, shift 3, 2048px; the official 8-step profile is a separate accelerated choice. Strengths: infographics, Chinese/English text, layouts. Proofread text, counts and placement; complex edits can affect untouched content.", "deepy_prompt_infos": '### Prompt\n\nDescribe subject, composition, style, palette and light. For infographics specify layout/panels, hierarchy and every exact string in quotes. For edits, state change and preservation constraints; number reference roles. Example: `Replace "CLOSED" with "OPEN", retaining the sign material and perspective.` Enhancement is optional; use it for short briefs needing design detail.', "preview_all_images": True, "prompt_enhancer_button_label": "Enhance", "prompt_enhancer_def": { "selection": ["T", "TI", "T1", "TI1"], "labels": { "T": "A General Image Prompt using existing Text Prompt", "TI": "A General Image Prompt using existing Text Prompt and {image_inputs}", "T1": "An Infographic Prompt using existing Text Prompt", "TI1": "An Infographic Prompt using existing Text Prompt and {image_inputs}", }, "default": "", }, "text_prompt_enhancer_instructions": SENSENOVA_GENERIC_PROMPT, "image_prompt_enhancer_instructions": SENSENOVA_GENERIC_REFERENCE_PROMPT, "text_prompt_enhancer_instructions1": SENSENOVA_INFOGRAPHIC_PROMPT, "image_prompt_enhancer_instructions1": SENSENOVA_INFOGRAPHIC_REFERENCE_PROMPT, "text_prompt_enhancer_max_tokens": 1024, "image_prompt_enhancer_max_tokens": 1024, "text_prompt_enhancer_max_tokens1": 1536, "image_prompt_enhancer_max_tokens1": 1536, "profiles_dir": [_PROFILE_FOLDER], "resolution_budget": "4096x4096", "skip_prompt_template": True, "vae_block_size": 32, } @staticmethod def query_supported_types(): return [_ARCHITECTURE] @staticmethod def query_family_maps(): return {}, {} @staticmethod def query_model_family(): return "sensenova" @staticmethod def query_family_infos(): return {"sensenova": (1170, "SenseNova")} @staticmethod def get_lora_dir(base_model_type): return _PROFILE_FOLDER @staticmethod def preview_latents(base_model_type, latents, meta): if not torch.is_tensor(latents) or latents.dim() != 4 or latents.shape[0] != 3: return None image = latents.detach().float().cpu().clamp(-1, 1) channels, frames, height, width = image.shape image = image.permute(0, 2, 1, 3).reshape(channels, height, frames * width) image = image.add(1).mul(127.5).clamp(0, 255).to(torch.uint8) preview = Image.fromarray(image.permute(1, 2, 0).numpy()) if preview.height > 0: scale = 200 / preview.height resampling_module = getattr(Image, "Resampling", Image) preview = preview.resize((max(1, int(round(preview.width * scale))), 200), resample=getattr(resampling_module, "BILINEAR", Image.BILINEAR)) return preview @staticmethod def query_model_files(computeList, base_model_type, model_def=None): return [{ "repoId": _PROJECT_REPO, "sourceFolderList": [_PROJECT_FOLDER], "fileList": [[ "added_tokens.json", "config.json", "merges.txt", "special_tokens_map.json", "tokenizer_config.json", "vocab.json", ]], }] @staticmethod def load_model( model_filename, model_type=None, base_model_type=None, model_def=None, quantizeTransformer=False, text_encoder_quantization=None, dtype=torch.bfloat16, VAE_dtype=torch.float32, mixed_precision_transformer=False, save_quantized=False, submodel_no_list=None, text_encoder_filename=None, **kwargs, ): from .sensenova_u1_main import model_factory processor = model_factory( model_filename=model_filename, model_type=model_type, base_model_type=base_model_type, dtype=dtype, save_quantized=save_quantized, ) return processor, {"transformer": processor.transformer} @staticmethod def update_default_settings(base_model_type, model_def, ui_defaults): ui_defaults.update({ "image_mode": 1, "num_inference_steps": 50, "guidance_scale": 4.0, "flow_shift": 3.0, "batch_size": 1, "video_prompt_type": "", })