license: apache-2.0
language:
- en
base_model: PrunaAI/P-H3-Image-Comfy
tags: - comfyui
- prompt-upsampling
- p-h3-image
- qwen3.5
- heretic
- int8
- w4a8
P-H3-Image Prompt Upsampler — Heretic
Heretic-modified versions of the P-H3-Image prompt upsampler, provided as ComfyUI single-file checkpoints in INT8 ConvRot and W4A8 formats. The upsampler is based on Qwen3.5-9B. It turns a short image request into a fuller prompt for P-H3-Image text-to-image generation or an edit brief for image editing. It is not the P-H3-Image diffusion model or its main image text encoder.
Files
| File | Format | Local file size |
|---|---|---|
p-h3-image_prompt_upsampler_heretic_int8_convrot.safetensors |
INT8 ConvRot | 11.46 GB |
p-h3-image_prompt_upsampler_heretic_w4a8.safetensors |
W4A8 | 7.46 GB |
Choose one file for a workflow. These are alternative quantizations of the Heretic-modified upsampler, not separate prompt stages. The listed sizes are the files on disk, in decimal GB; they are not measured peak VRAM requirements.
What was changed
The prompt upsampler was modified with Heretic, using the default refusal-suppression configuration and the ARA (Arbitrary-Rank Ablation) modifier. The selected result is Trial 424, from a 450-trial optimization run.
Heretic evaluation results
| Measure | Original model | Selected Trial 424 |
|---|---|---|
| Refusals on Heretic's harmful-behavior evaluation prompts | 95/100 | 6/100 |
| KL divergence on Heretic's harmless Alpaca prompts | 0 (baseline) | 0.0069 |
Heretic's log reports 100 prompts for each evaluation measure and 400 prompts per category for optimization. These are Heretic's internal selection metrics, not independent safety, prompt-quality, or image-quality benchmarks. The results describe this evaluation run and do not guarantee how the model will respond to other prompts. The original source includes 450 optimization trials; Trial 424 was selected from the resulting Pareto-optimal options.
Use in ComfyUI
- Put either safetensors file in
ComfyUI/models/text_encoders/and refresh ComfyUI's model list. - Open a current P-H3-Image ComfyUI workflow.
- In the Load CLIP node used for the prompt upsampler, select this file and keep the loader type set to
minimax. This is separate from the Load CLIP node for P-H3-Image's main text encoder. - Keep the upsampler's Generate Text output connected to the prompt input of the P-H3-Image conditioning node, as in the upstream workflow.
For text-to-image, the upstream workflow supplies a short request with an aspect ratio, for example:
Aspect ratio: 1:1
Request: A candid street photograph of a rainy market at dusk.
For image editing, use the upstream edit workflow's image inputs and request format. The upsampler does not generate images by itself; the P-H3-Image diffusion model, main text encoder, VAE, and conditioning workflow are separate downloads.
Limitations
- Heretic's default settings aim to reduce refusals; they do not guarantee that every request will be accepted or that an expanded prompt will preserve every detail of the request.
- Prompt expansion can add or alter scene details. Review the generated prompt before relying on exact text, identity, or edit instructions.
- Quantization can change generated text. Neither file has a published comparison against the original BF16 upsampler in this release.
- No performance, VRAM, safety, or image-quality benchmark is claimed for these modified files.
Provenance and license
The source component is the prompt upsampler in PrunaAI/P-H3-Image-Comfy, which is based on Qwen3.5-9B. Pruna AI's model card identifies the prompt upsampler as Apache 2.0 licensed. See the upstream Apache license and NOTICE, and include those files with this upload.
The P-H3-Image diffusion model has separate MiniMax H3 Community License terms. This card's Apache 2.0 metadata applies to the prompt upsampler files in this repository, not to the full image model.