nifleisch commited on
Commit
fd9b9e1
·
verified ·
1 Parent(s): 62f4d32

Add model card

Browse files

Model card with quickstart, file and VRAM guide, recommended resolutions, prompt upsampler and editing guidance, and figures.

.gitattributes CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ assets/multi_image.jpg filter=lfs diff=lfs merge=lfs -text
37
+ assets/p-h3-image-comfy.jpg filter=lfs diff=lfs merge=lfs -text
38
+ assets/quantization.jpg filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -21,47 +21,252 @@ tags:
21
  - minimax-h3
22
  - pruna-ai
23
  ---
 
 
 
 
 
24
 
25
- # p-h3-image for ComfyUI
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
26
 
27
- p-h3-image, based on MiniMax-H3, makes images from text and edits images, with up to 9 input images. This repository
28
- has its ComfyUI files.
29
 
30
- ## Use it
31
 
32
- 1. Install the nodes (ComfyUI 0.43 or newer):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33
 
34
  ```bash
35
  git clone https://github.com/PrunaAI/comfyui-p-h3-image ComfyUI/custom_nodes/comfyui-p-h3-image
36
  ```
37
 
38
- 2. Download the files into `ComfyUI/models`:
 
39
 
40
  ```bash
41
  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
42
  ```
43
 
44
- Less VRAM: take the `w4a8` diffusion model and text encoder and the `int8_convrot` VAE instead.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45
 
46
- 3. Open a workflow from `workflows/` (also under Templates in ComfyUI), type a short request and run.
 
 
47
 
48
- Keep the prompt upsampler in the graph: it turns a short request, in any language, into the full prompt p-h3-image
49
- needs. The workflows default to 8 megapixels and 8 steps.
50
 
51
- Already have MiniMax-H3 (FL2VA) in ComfyUI? Add `loras/p-h3-image_lora_bf16.safetensors` and use the `lora` workflows.
 
 
52
 
53
- ## Files
 
 
 
 
 
 
 
 
54
 
55
- ```
56
- ComfyUI/models/
57
- ├── diffusion_models/ p-h3-image_{bf16,int8_convrot,w4a8}.safetensors
58
- ├── text_encoders/ p-h3-image_text_encoder_{bf16,int8_convrot,w4a8}.safetensors
59
- │ p-h3-image_prompt_upsampler_{bf16,int8_convrot,w4a8}.safetensors
60
- ├── vae/ p-h3-image_vae_{bf16,int8_convrot}.safetensors
61
- └── loras/ p-h3-image_lora_bf16.safetensors
62
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63
 
64
  ## License
65
 
66
- The p-h3-image weights are under the MiniMax H3 Community License ([LICENSE](LICENSE)), the prompt upsampler under
67
- Apache 2.0 ([LICENSE-APACHE](LICENSE-APACHE)). See [NOTICE](NOTICE).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ The nodes are in [PrunaAI/comfyui-p-h3-image](https://github.com/PrunaAI/comfyui-p-h3-image). The diffusers pipeline is
76
+ in [PrunaAI/p-h3-image](https://huggingface.co/PrunaAI/p-h3-image).
77
+
78
+ ## Quickstart
79
+
80
+ 1. **Install the nodes.** ComfyUI 0.43 or newer, no extra dependencies.
81
 
82
  ```bash
83
  git clone https://github.com/PrunaAI/comfyui-p-h3-image ComfyUI/custom_nodes/comfyui-p-h3-image
84
  ```
85
 
86
+ 2. **Download one file per component** into `ComfyUI/models`. The bf16 set, a 55 GB download plus the prompt upsampler,
87
+ peaks at about 56 GB of VRAM:
88
 
89
  ```bash
90
  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
91
  ```
92
 
93
+ The smallest set, a 20 GB download plus the prompt upsampler, peaks at about 21 GB of VRAM:
94
+
95
+ ```bash
96
+ 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
97
+ ```
98
+
99
+ 3. **Open a workflow** from `workflows/` (also under Templates in ComfyUI), type a short request and run. The
100
+ workflows render at 8 MP in 8 steps through the prompt upsampler.
101
+
102
+ ## Pick your files
103
+
104
+ Pick one file per component. Files of different precisions can be mixed freely.
105
+
106
+ | Component | Folder | bf16 | int8 (`int8_convrot`) | w4a8 |
107
+ |---|---|:---:|:---:|:---:|
108
+ | Diffusion model | `diffusion_models/` | 40.6 GB | 21.3 GB | 13.5 GB |
109
+ | Text encoder | `text_encoders/` | 9.1 GB | 5.1 GB | 3.6 GB |
110
+ | Prompt upsampler | `text_encoders/` | 19.3 GB | 9.7 GB | 6.7 GB |
111
+ | VAE | `vae/` | 5.6 GB | 3.1 GB | – |
112
+ | LoRA for MiniMax-H3 | `loras/` | 2.0 GB | – | – |
113
+
114
+ | Set (model, text encoder, VAE) | Download | Peak VRAM at 8 MP |
115
+ |---|:---:|:---:|
116
+ | bf16 | 55 GB | 56 GB |
117
+ | int8 | 30 GB | 31 GB |
118
+ | w4a8, int8 VAE | **20 GB** | **21 GB** |
119
+
120
+ <p align="center">
121
+ <img src="assets/weights.png" width="100%">
122
+ </p>
123
+
124
+ *Weights on disk. MiniMax-H3 counts its transformer, text encoder and video VAE.*
125
+
126
+ The int8 set stays very close to bf16, often almost pixel for pixel. The w4a8 set keeps the quality but changes more
127
+ of the details and sometimes the composition. On a 48 GB L40S, where the bf16 set has to offload, the w4a8 set renders a
128
+ 2048 × 2048 image in 6.3 s instead of 19 s.
129
+
130
+ <p align="center">
131
+ <img src="assets/quantization.jpg" width="100%">
132
+ </p>
133
+
134
+ *The same prompts and seed in bf16, int8 and w4a8, at 8 MP. The headers give the download size and the peak VRAM at
135
+ 8 MP.*
136
+
137
+ The quantized files use ComfyUI's native formats, need no calibration data and run on comfy-kitchen's INT8 kernels, for
138
+ NVIDIA Ampere or newer and AMD RDNA3/4.
139
+
140
+ - **int8 (`int8_convrot`).** Weights are rotated with grouped Hadamard transforms to spread outliers, then stored as
141
+ INT8 with per-row scales. Activations are quantized to INT8 on the fly.
142
+ - **w4a8.** The same rotation, then 4-bit weights on a 16-level codebook with an FP8 scale for every 16 inputs, and
143
+ INT8 activations. The layers most sensitive to 4-bit weights stay in int8.
144
+
145
+ ## Workflows
146
+
147
+ | Workflow | What it does |
148
+ |---|---|
149
+ | `p-h3-image_text_to_image.json` | Text-to-image at 8 MP from a short request |
150
+ | `p-h3-image_edit.json` | Edits an image, with optional reference images |
151
+ | `p-h3-image_lora_text_to_image.json` | Text-to-image with the LoRA on your own MiniMax-H3 |
152
+ | `p-h3-image_lora_edit.json` | Editing with the LoRA on your own MiniMax-H3 |
153
 
154
+ The pack adds loaders for the model, text encoder, VAE and LoRA, an Empty Latent node, an Edit node, a Sigmas node
155
+ with the 8-step grid, and two prompt upsampler nodes, one for text-to-image and one for edits. Sampling uses ComfyUI's
156
+ Basic Guider, `euler` and Sampler Custom Advanced, with no CFG or negative prompt.
157
 
158
+ ## Recommended resolutions
 
159
 
160
+ <p align="center">
161
+ <img src="assets/resolutions.png" width="100%">
162
+ </p>
163
 
164
+ | Aspect ratio | 8 MP (width × height) | 4 MP, for edits |
165
+ |:---:|:---:|:---:|
166
+ | 1:1 | **2912 × 2912** | 2048 × 2048 |
167
+ | 4:3 | **3360 × 2496** | 2368 × 1760 |
168
+ | 3:4 | **2496 × 3360** | 1760 × 2368 |
169
+ | 3:2 | **3552 × 2368** | 2496 × 1664 |
170
+ | 2:3 | **2368 × 3552** | 1664 × 2496 |
171
+ | 16:9 | **3872 × 2176** | 2720 × 1536 |
172
+ | 9:16 | **2176 × 3872** | 1536 × 2720 |
173
 
174
+ Text-to-image works best at 8 MP. The text-to-image upsampler node returns the 8 MP width and height for the aspect
175
+ ratio you pick. Edits work best at 4 or 8 MP. The Edit node scales `image_1`, the image being edited, up to at least
176
+ 4 MP and generates the edit at its size, unless `width` and `height` are set. Reference images keep their size between
177
+ 1 and 8 MP.
178
+
179
+ ## Prompt upsampler
180
+
181
+ p-h3-image is sensitive to prompting and renders full, specific prompts much better than short requests. The prompt
182
+ upsampler, based on Qwen3.5-9B, turns a one-line request into that prompt, and all example workflows use it. It thinks
183
+ about the request first and then writes the prompt.
184
+
185
+ - **p-h3-image Prompt Upsampler.** For text-to-image. Takes the request and an aspect ratio, and returns the prompt for
186
+ CLIP Text Encode and the 8 MP width and height for p-h3-image Empty Latent.
187
+ - **p-h3-image Edit Prompt Upsampler.** For edits. Takes the request and 1 to 8 images, the image to edit first, and
188
+ returns the edit prompt for the p-h3-image Edit node. Give the Edit node the same images in the same order.
189
+
190
+ Load it with Load CLIP, type `minimax`. It runs before p-h3-image, and ComfyUI moves it out of VRAM when p-h3-image
191
+ needs the memory. Its output is cached, so it only runs again when the request or the images change. Decoding is
192
+ greedy, so a request always gives the same prompt. Turn on `sampling` for a different prompt per seed.
193
+
194
+ | Upsampler file | Size | Peak VRAM | Time per prompt, L40S |
195
+ |---|:---:|:---:|:---:|
196
+ | `p-h3-image_prompt_upsampler_bf16` (recommended) | 19.3 GB | about 21 GB | about 7 s |
197
+ | `p-h3-image_prompt_upsampler_int8_convrot` | 9.7 GB | about 12 GB | about 13 s |
198
+ | `p-h3-image_prompt_upsampler_w4a8` | 6.7 GB | about 9 GB | about 11 s |
199
+
200
+ The w4a8 prompts differ most from bf16. Use the smaller files only when 21 GB does not fit.
201
+
202
+ ## Editing
203
+
204
+ Connect the image to edit as `image_1` and up to 7 references after it, in the same order on the Edit Prompt Upsampler
205
+ and the Edit node. The request can name the images however you like, such as "the hat from the second photo".
206
+
207
+ <p align="center">
208
+ <img src="assets/multi_image.jpg" width="100%">
209
+ </p>
210
+
211
+ *Top, eight public-domain portraits and the dinner party p-h3-image composed from them, from a one-sentence request.
212
+ Bottom, eight scientists at a poker table, seven presidents wading through the swamp with an astronaut, and characters
213
+ from our own renders in the same swamp.*
214
+
215
+ ## Already have MiniMax-H3?
216
+
217
+ 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
218
+ transformer without a second large download. Load `minimax_h3_fl2va_bf16.safetensors` from
219
+ [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) (the pruned and quantized FL2VA files work too),
220
+ apply the LoRA with the p-h3-image LoRA Loader at strength 1.0, and start from the `lora` workflows.
221
+
222
+ ## Limitations
223
+
224
+ - **It is sensitive to prompting.** Keep the prompt upsampler in the graph. Hype words and very long prompts make
225
+ images worse, and very short prompts give weak results without it.
226
+ - **Hands can break, and small people in the background can come out distorted.** Rendering at 8 MP helps.
227
+ - **Text is not fully reliable.** Small or dense text turns into pseudo-text, and longer passages sometimes get single
228
+ letters wrong. Short quoted strings render well.
229
+ - **Quantized files are not identical to bf16.** int8 stays very close, w4a8 changes more of the details.
230
 
231
  ## License
232
 
233
+ The p-h3-image weights are based on MiniMax-H3 and released under the [MiniMax H3 Community License](LICENSE). Please
234
+ read it before use. Among other terms, it excludes the European Union, the United Kingdom, the Republic of Korea and
235
+ the United States unless you have a separate license from MiniMax, and products built on it must show "Powered by
236
+ MiniMax H3". The text encoder's Qwen3-VL-4B parts and the prompt upsampler, based on Qwen3.5-9B, are under the
237
+ [Apache License 2.0](LICENSE-APACHE). See [NOTICE](NOTICE) for details. The ComfyUI nodes are MIT-licensed.
238
+
239
+ ## What's next?
240
+
241
+ - Compress your own models with [Pruna](https://github.com/PrunaAI/pruna) and give us a ⭐️ for more efficiency!
242
+ - 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).
243
+
244
+ <style>
245
+ .model-button {
246
+ display: inline-flex;
247
+ flex-direction: row;
248
+ justify-content: center;
249
+ align-items: center;
250
+ gap: 8px;
251
+
252
+ padding: 8px 20px;
253
+ border: none;
254
+ border-radius: 8px;
255
+
256
+ background: #9334e9;
257
+ color: #ffffff;
258
+
259
+ font-size: 14px;
260
+ font-weight: 400;
261
+ line-height: 1;
262
+ text-decoration: none;
263
+ white-space: nowrap;
264
+ cursor: pointer;
265
+
266
+ box-sizing: border-box;
267
+ overflow: visible;
268
+ opacity: 1;
269
+ }
270
+ </style>
271
+
272
+ <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>
assets/multi_image.jpg ADDED

Git LFS Details

  • SHA256: ee0c9775b222d2be273fe3303fec57ad61c44309303f72005c768ca29e0cd64a
  • Pointer size: 131 Bytes
  • Size of remote file: 761 kB
assets/p-h3-image-comfy.jpg ADDED

Git LFS Details

  • SHA256: 6d2a5037b5116b6841b89c964cafbb23ff3bd8e3ad84cd9db2ff51b83659ab2b
  • Pointer size: 132 Bytes
  • Size of remote file: 1.02 MB
assets/quantization.jpg ADDED

Git LFS Details

  • SHA256: d4d397138902a8c3cecbbdfab8e5fcc9f94123ab4490b2cfcd519f4eaa354e5e
  • Pointer size: 131 Bytes
  • Size of remote file: 790 kB
assets/resolutions.png ADDED
assets/weights.png ADDED