README.md
17.4 KB · 282 lines · markdown Raw
1 ---
2 license: other
3 license_name: minimax-h3-community-license-agreement
4 license_link: LICENSE
5 base_model: PrunaAI/p-h3-image
6 base_model_relation: quantized
7 pipeline_tag: text-to-image
8 language:
9 - en
10 tags:
11 - comfyui
12 - diffusion-single-file
13 - text-to-image
14 - image-to-image
15 - image-generation
16 - image-editing
17 - prompt-upsampling
18 - lora
19 - int8
20 - 4-bit
21 - minimax-h3
22 - pruna-ai
23 ---
24 <!-- header start -->
25 <p align="center">
26 <img src="assets/p-h3-image-comfy.jpg" width="800">
27 </p>
28 <!-- header end -->
29
30 <p align="center">
31 <a href="https://github.com/PrunaAI/pruna">
32 <img src="https://img.shields.io/badge/GitHub-PrunaAI-9334E9?style=plastic&logo=github&logoColor=white" alt="GitHub">
33 </a>
34 &nbsp;
35 <a href="https://twitter.com/PrunaAI">
36 <img src="https://img.shields.io/badge/Twitter%2FX-@PrunaAI-9334E9?style=plastic&logo=x&logoColor=white" alt="Twitter/X">
37 </a>
38 &nbsp;
39 <a href="https://www.linkedin.com/company/pruna-ai">
40 <img src="https://img.shields.io/badge/LinkedIn-PrunaAI-9334E9?style=plastic&logo=linkedin&logoColor=white" alt="LinkedIn">
41 </a>
42 &nbsp;
43 <a href="https://discord.com/invite/JFQmtFKCjd">
44 <img src="https://img.shields.io/badge/Discord-Join%20us-9334E9?style=plastic&logo=discord&logoColor=white" alt="Discord">
45 </a>
46 &nbsp;
47 <a href="https://dashboard.pruna.ai/login?utm_source=huggingface&utm_medium=org_card&utm_campaign=hf_traffic">
48 <img src="https://img.shields.io/badge/Performance%20Models-Try%20them%20now-9334E9?style=plastic&logo=data:image/svg%2bxml;base64,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" alt="Performance Models">
49 </a>
50 </p>
51 <div align="center">
52
53 <h1 style="color: #9334E9;">P-H3-Image</h1>
54
55 <h2>ComfyUI files for P-H3-Image, based on MiniMax-H3</h2>
56
57 <h3>
58 <span style="color: #9334E9;">bf16, int8 and w4a8</span>
59 &nbsp;·&nbsp;
60 <span style="color: #9334E9;">From 21 GB of VRAM</span>
61 &nbsp;·&nbsp;
62 <span style="color: #9334E9;">8 steps</span>
63 &nbsp;·&nbsp;
64 Text-to-image and editing with up to 8 images
65 </h3>
66
67 </div>
68
69 This repository holds the ComfyUI files for **P-H3-Image**, an image model for text-to-image generation and image
70 editing based on [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3). It renders 8-megapixel images in 8 steps
71 without CFG and edits with up to 8 input images. Every component comes in bf16 and in ComfyUI's native int8 and 4-bit
72 formats. The smallest set is a 20 GB download that peaks at 21 GB of VRAM at 8 MP, against 143 GB of weights for
73 MiniMax-H3.
74
75 ComfyUI supports P-H3-Image natively, no custom nodes needed. The diffusers pipeline is
76 in [PrunaAI/P-H3-Image](https://huggingface.co/PrunaAI/P-H3-Image). Try it in the [Hugging Face Spaces](https://huggingface.co/spaces/PrunaAI/P-H3-Image). Read the full blog post [here](https://pruna.ai/blog/oss-p-h3-image-launch).
77
78 ## Quickstart
79
80 1. **Update ComfyUI.** P-H3-Image support is on ComfyUI's master branch and ships with the next release. Until then,
81 update to the latest ComfyUI from GitHub (nightly). The desktop app and Comfy Cloud follow stable releases and get it
82 with the next one.
83
84 ```bash
85 cd ComfyUI && git pull
86 ```
87
88 2. **Download one file per component** into `ComfyUI/models`. The bf16 set, a 55 GB download plus the prompt upsampler,
89 peaks at about 56 GB of VRAM:
90
91 ```bash
92 hf download PrunaAI/P-H3-Image-comfy --include "diffusion_models/p-h3-image_bf16.safetensors" "text_encoders/p-h3-image_text_encoder_bf16.safetensors" "text_encoders/p-h3-image_prompt_upsampler_bf16.safetensors" "vae/p-h3-image_vae_bf16.safetensors" --local-dir ComfyUI/models
93 ```
94
95 The smallest set, a 20 GB download plus the prompt upsampler, peaks at about 21 GB of VRAM:
96
97 ```bash
98 hf download PrunaAI/P-H3-Image-Comfy --include "diffusion_models/p-h3-image_w4a8.safetensors" "text_encoders/p-h3-image_text_encoder_w4a8.safetensors" "text_encoders/p-h3-image_prompt_upsampler_bf16.safetensors" "vae/p-h3-image_vae_int8_convrot.safetensors" --local-dir ComfyUI/models
99 ```
100
101 3. **Open a workflow** from `workflows/`, type a short request and run. The
102 workflows render at 8 MP in 8 steps through the prompt upsampler.
103
104 ## Pick your files
105
106 Pick one file per component. Files of different precisions can be mixed freely.
107
108 | Component | Folder | bf16 | int8 (`int8_convrot`) | w4a8 |
109 |---|---|:---:|:---:|:---:|
110 | Diffusion model | `diffusion_models/` | 40.5 GB | 21.2 GB | 13.4 GB |
111 | Text encoder | `text_encoders/` | 9.1 GB | 5.1 GB | 3.6 GB |
112 | Prompt upsampler | `text_encoders/` | 19.3 GB | 9.7 GB | 6.7 GB |
113 | VAE | `vae/` | 5.6 GB | 3.1 GB | – |
114 | LoRA for MiniMax-H3 | `loras/` | 2.0 GB | – | – |
115
116 | Set (model, text encoder, VAE) | Download | Peak VRAM at 8 MP |
117 |---|:---:|:---:|
118 | bf16 | 55 GB | 56 GB |
119 | int8 | 30 GB | 31 GB |
120 | w4a8, int8 VAE | **20 GB** | **21 GB** |
121
122 <p align="center">
123 <img src="assets/weights.png" width="100%">
124 </p>
125
126 *Weights on disk. MiniMax-H3 counts its transformer, text encoder and video VAE.*
127
128 The int8 set stays very close to bf16, often almost pixel for pixel. The w4a8 set keeps the quality but changes more
129 of the details and sometimes the composition. On a 48 GB L40S, where the bf16 set has to offload, the w4a8 set renders a
130 2048 × 2048 image in 6.3 s instead of 19 s.
131
132 <p align="center">
133 <img src="assets/quantization.jpg" width="100%">
134 </p>
135
136 *The same prompts and seed in bf16, int8 and w4a8, at 8 MP. The headers give the download size and the peak VRAM at
137 8 MP.*
138
139 The quantized files use ComfyUI's native formats, need no calibration data and run on comfy-kitchen's INT8 kernels, for
140 NVIDIA Ampere or newer and AMD RDNA3/4.
141
142 - **int8 (`int8_convrot`).** Weights are rotated with grouped Hadamard transforms to spread outliers, then stored as
143 INT8 with per-row scales. Activations are quantized to INT8 on the fly.
144 - **w4a8.** The same rotation, then 4-bit weights on a 16-level codebook with an FP8 scale for every 16 inputs, and
145 INT8 activations. The layers most sensitive to 4-bit weights stay in int8.
146
147 ## Workflows
148
149 | Workflow | What it does |
150 |---|---|
151 | `p-h3-image_text_to_image.json` | Text-to-image at 8 MP from a short request |
152 | `p-h3-image_edit.json` | Edits an image, with optional reference images |
153 | `p-h3-image_lora_text_to_image.json` | Text-to-image with the LoRA on your own MiniMax-H3 |
154 | `p-h3-image_lora_edit.json` | Editing with the LoRA on your own MiniMax-H3 |
155
156 The workflows use only ComfyUI's own nodes. Load Diffusion Model loads the model, followed by ModelSamplingMiniMaxH3 at
157 1.0 for both shifts. Load CLIP (type `minimax`) loads the text encoder, Load VAE the VAE, and p-h3-image Conditioning
158 turns the prompt and the images into the conditioning and the latent. The prompt upsampler runs on Generate Text, and
159 sampling on ManualSigmas with the 8-step grid, Basic Guider, `euler` and Sampler Custom Advanced, with no CFG or
160 negative prompt. The LoRA loads with ComfyUI's Load LoRA (Bypass, Model Only).
161
162 ## Recommended resolutions
163
164 <p align="center">
165 <img src="assets/resolutions.png" width="100%">
166 </p>
167
168 | Aspect ratio | 8 MP (width × height) | 4 MP, for edits |
169 |:---:|:---:|:---:|
170 | 1:1 | **2912 × 2912** | 2048 × 2048 |
171 | 4:3 | **3360 × 2496** | 2368 × 1760 |
172 | 3:4 | **2496 × 3360** | 1760 × 2368 |
173 | 3:2 | **3552 × 2368** | 2496 × 1664 |
174 | 2:3 | **2368 × 3552** | 1664 × 2496 |
175 | 16:9 | **3872 × 2176** | 2720 × 1536 |
176 | 9:16 | **2176 × 3872** | 1536 × 2720 |
177
178 Text-to-image works best at 8 MP. In the workflows, Resolution Selector at 8 megapixels and a multiple of 32 gives
179 these sizes. Edits work best at 4 or 8 MP. In the edit workflows, Scale Image to Total Pixels brings `image_1`, the
180 image being edited, to 4 MP, and p-h3-image Conditioning generates the edit at its size unless `width` and `height` are
181 set. Reference images above about 8 MP are scaled down.
182
183 ## Prompt upsampler
184
185 P-H3-Image is sensitive to prompting and renders full, specific prompts much better than short requests. The prompt
186 upsampler, based on Qwen3.5-9B, turns a one-line request into that prompt, and all example workflows use it. It thinks
187 about the request first and then writes the prompt.
188
189 It runs on ComfyUI's own nodes, the way FLUX.3 Klein's templates run theirs: Load CLIP (type `minimax`) and Generate
190 Text, with greedy decoding, so a request always gives the same prompt.
191
192 - **Text-to-image.** Format Text turns the request into the upsampler's input, `Aspect ratio: 1:1` and `Request: ...`
193 on two lines, and Generate Text runs with `thinking` on. Keep the aspect ratio the same as in Resolution Selector.
194 - **Edits.** Format Text writes the upsampler's input with a `<Picture N>` block for every image, and the images reach
195 Generate Text through Batch Images: `image_1` at 1 MP, the references padded to its size. The result is the edit
196 prompt for p-h3-image Conditioning.
197 - **Use prompt upsampler**, an If/Else switch, set to false sends the request to the model as it is.
198
199 The upsampler runs before P-H3-Image, and ComfyUI moves it out of VRAM when P-H3-Image needs the memory. Its output is
200 cached, so it only runs again when the request or the images change.
201
202 | Upsampler file | Size | Peak VRAM | Time per prompt, L40S |
203 |---|:---:|:---:|:---:|
204 | `p-h3-image_prompt_upsampler_bf16` (recommended) | 19.3 GB | about 21 GB | about 7 s |
205 | `p-h3-image_prompt_upsampler_int8_convrot` | 9.7 GB | about 12 GB | about 13 s |
206 | `p-h3-image_prompt_upsampler_w4a8` | 6.7 GB | about 9 GB | about 11 s |
207
208 The w4a8 prompts differ most from bf16. Use the smaller files only when 21 GB does not fit.
209
210 ## Editing
211
212 Connect the image to edit as `image_1` and up to 7 references after it, in the same order in Batch Images and on
213 p-h3-image Conditioning, with a `<Picture N>` block in Upsampler input for every image. The request can name the images
214 however you like, such as "the hat from the second photo".
215
216 <p align="center">
217 <img src="assets/multi_image.jpg" width="100%">
218 </p>
219
220 *Top, eight public-domain portraits and the dinner party P-H3-Image composed from them, from a one-sentence request.
221 Bottom, eight scientists at a poker table, seven presidents wading through the swamp with an astronaut, and characters
222 from our own renders in the same swamp.*
223
224 ## Already have MiniMax-H3?
225
226 If you run MiniMax-H3 in ComfyUI, the 2 GB LoRA `loras/p-h3-image_lora_bf16.safetensors` runs P-H3-Image on its FL2VA
227 transformer without a second large download. Load `minimax_h3_fl2va_bf16.safetensors` from
228 [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) (the pruned and quantized FL2VA files work too),
229 apply the LoRA with Load LoRA (Bypass, Model Only) at strength 1.0, and start from the `lora` workflows.
230
231 ## Limitations
232
233 - **It is sensitive to prompting.** Keep the prompt upsampler in the graph. Hype words and very long prompts make
234 images worse, and very short prompts give weak results without it.
235 - **Hands can break, and small people in the background can come out distorted.** Rendering at 8 MP helps.
236 - **Text is not fully reliable.** Small or dense text turns into pseudo-text, and longer passages sometimes get single
237 letters wrong. Short quoted strings render well.
238 - **Quantized files are not identical to bf16.** int8 stays very close, w4a8 changes more of the details.
239
240 ## License
241
242 The P-H3-Image weights are based on MiniMax-H3 and released under the [MiniMax H3 Community License](LICENSE). Please
243 read it before use. Among other terms, it excludes the European Union, the United Kingdom, the Republic of Korea and
244 the United States unless you have a separate license from MiniMax, and products built on it must show "Powered by
245 MiniMax H3". The text encoder's Qwen3-VL-4B parts and the prompt upsampler, based on Qwen3.5-9B, are under the
246 [Apache License 2.0](LICENSE-APACHE). See [NOTICE](NOTICE) for details.
247
248 ## What's next?
249
250 - Compress your own models with [Pruna](https://github.com/PrunaAI/pruna) and give us a ⭐️ for more efficiency!
251 - Want to use our optimized image models right away? Check [P-Image-Ideogram](https://www.pruna.ai/p-image-ideogram) and [P-Image-Edit](https://www.pruna.ai/p-image-edit).
252
253 <style>
254 .model-button {
255 display: inline-flex;
256 flex-direction: row;
257 justify-content: center;
258 align-items: center;
259 gap: 8px;
260
261 padding: 8px 20px;
262 border: none;
263 border-radius: 8px;
264
265 background: #9334e9;
266 color: #ffffff;
267
268 font-size: 14px;
269 font-weight: 400;
270 line-height: 1;
271 text-decoration: none;
272 white-space: nowrap;
273 cursor: pointer;
274
275 box-sizing: border-box;
276 overflow: visible;
277 opacity: 1;
278 }
279 </style>
280
281 <a href="https://dashboard.pruna.ai/login?utm_source=huggingface&utm_medium=model_card&utm_campaign=hf_traffic" class="model-button">Try our models</a>
282