P-H3-Image (A 800 megapixel text-to-image / image editing model based on MiniMax-H3, using 8 steps)

An image model fine-tuned by Pruna AI on MiniMax-H3: Text-to-image and image editing share a single pipeline, requiring 8 steps without CFG or negative prompts, directly generating images at 800 megapixels (8 MP). For image editing, up to 8 images can be provided; the 1th image is the one to be edited, while the others serve as references (referred to in the prompt as Picture 2, Picture 3, etc.). It includes a built-in prompt expansion model based on Qwen3.5-9B, enabled by default, which rewrites brief requests into complete prompts that it handles best. When editing images, it formats instructions as "what to change and what to preserve." The repository was created on 10/27, with update notes added on 11/3. It has not been featured in the last 7 issues.

An image model fine-tuned by Pruna AI on MiniMax-H3: Text-to-image and image editing share a single pipeline, requiring 8 steps without CFG or negative prompts, directly generating images at 800 megapixels (8 MP). For image editing, up to 8 images can be provided; the 1th image is the one to be edited, while the others serve as references (referred to in the prompt as Picture 2, Picture 3, etc.). It includes a built-in prompt expansion model based on Qwen3.5-9B, enabled by default, which rewrites brief requests into complete prompts that it handles best. When editing images, it formats instructions as "what to change and what to preserve." The repository was created on 10/27, with update notes added on 11/3. It has not been featured in the last 7 issues.

Rising score: 106 · Likes: 109 · Downloads: 287

Reasons

  • Addresses the need for "generating high-resolution images from a single sentence and editing images without redrawing the entire picture": The author states it excels at photorealistic photography, cinematic and magazine-style visuals, and graphic design featuring short quoted text; the author's comparison images show that directly inputting short prompts often results in redrawing the entire image, whereas expanding them into editorial prompts modifies only the specified parts.
  • Officially, rendering a single image at B200, bf16, and 8 MP takes approximately 9 seconds with peak VRAM usage of 83 GB; enabling prompt expansion adds 6–11 seconds and increases peak VRAM to approximately 101 GB. The architecture consists of a 33.1B parameter transformer, a 4B text encoder, a VAE, and an expansion model, totaling 100 GB in bf16.
  • The author also released a ComfyUI version (PrunaAI/P-H3-Image-Comfy) that uses only native ComfyUI nodes, offering three precision options for each component—bf16, int8, and w4a8—which can be mixed and matched: the smallest configuration requires a 20 GB download and peaks at approximately 21 GB of VRAM at 8 MP; the author claims int8 is nearly pixel-identical to bf16, while w4a8 alters details and occasionally changes composition. On an L40S with 48 GB of VRAM, w4a8 generates 2048×2048 images in approximately 6.3 seconds (bf16 requires offloading and takes approximately 19 seconds). Additionally, a 2.0 GB LoRA is provided, which can be attached to the author's MiniMax-H3.

Usage

  • ComfyUI: First, run cd ComfyUI && git pull to update to the latest master branch (stable releases, desktop versions, and Comfy Cloud will need to wait for the next release). Download one file for each component from PrunaAI/P-H3-Image-Comfy into the corresponding directories under ComfyUI/models. Open p-h3-image_text_to_image.json or p-h3-image_edit.json from the workflows/ directory, enter a one-sentence prompt, and run. The workflow uses Euler sampling, 8 steps, and no CFG; set the Load CLIP type to minimax.
  • diffusers: Run pip install git+https://github.com/huggingface/diffusers，再用 and use ModularPipeline.from_pretrained('PrunaAI/P-H3-Image') with load_components(dtype=torch.bfloat16). Passing an image enables image editing, while passing a list of images enables multi-image reference. For 80 GB GPUs, use enable_auto_cpu_offload from ComponentsManager; the official documentation states that peak memory usage for 8 MP is approximately 76 GB.
  • Select resolution based on the 8 MP size table for your GPU (default is 2912×2912; use 3872×2176 for 16:9). For image editing, use 4 or 8 MP. When writing prompts manually, start the first sentence by clearly stating "what this is" (e.g., "a photo taken by a street photographer"). Enclose any text that should appear in the image in quotation marks, character by character, and end the prompt with "no other text". Avoid words like "stunning", "8k", or "masterpiece" (the author notes these can result in an oily, oversaturated look). If you prefer not to download anything, you can try it first on the Space PrunaAI/P-H3-Image.

Risks

  • License: MiniMax H3 Community License. Without separate authorization from MiniMax, it does not cover the EU, UK, South Korea, and the US. Products based on it must be labeled "Powered by MiniMax H3." Only the Qwen3-VL-4B text encoder component and the expansion model are licensed under Apache-2.0.
  • Limitations listed by the author: Sensitive to prompts; exaggerated adjectives or overly long prompts degrade quality. Hands and small figures in backgrounds may distort. Small text and dense body text often become pseudo-text. Errors in faces, hands, and details increase at 2–4 MP. It generates only single images, not video or audio. This column has not reviewed example images.
  • High barrier to entry: The original diffusers version requires approximately 101 GB VRAM for 8 MP (approximately 83 GB with expansion disabled). The ComfyUI version is currently only available in the development branch; int8 / w4a8 depend on INT8 kernels from comfy-kitchen, requiring NVIDIA Ampere or newer, or AMD RDNA3/4. Speed and quality comparisons are based solely on the author's data.

Open

Hugging Face Trending

Explore the latest AI Radar updates