realrebelai/P-H3-Image_GGUFs

P-H3-Image — GGUF, W4A8, W6A8 & INT8 ConvRot

Community quantizations of P-H3-Image for memory-conscious local image generation.

Created by Rebel AI. These are unofficial conversions, not releases from PrunaAI or the original model developers.

Available quantizations

Format	Type	Intended trade-off	Notes
Q8_0	GGUF	Highest precision in this GGUF ladder	Transformer weight tensors quantized to Q8_0; other tensors retained at F16/F32
Q6_K	GGUF	Quality-oriented K-quant	Transformer weight tensors quantized to Q6_K; other tensors retained at F16/F32
Q5_K_M	GGUF	Middle ground	Mixed Q5_K/Q6_K policy for selected transformer weights; not a universal llama.cpp preset
Q4_K_M	GGUF	Smaller GGUF option	Mixed Q4_K/Q6_K policy for selected transformer weights; not a universal llama.cpp preset
W6A8	Safetensors	Higher weight precision than W4A8	6-bit weights / 8-bit activation-oriented quantization; ConvRot group size 256
W4A8	Safetensors	More aggressive weight compression	4-bit weights / 8-bit activation-oriented quantization; loader support required
INT8 ConvRot	Safetensors	8-bit weight representation	ConvRot-based quantization; loader support required

Important: W4A8 and W6A8 are not GGUF formats. They are distinct quantized safetensors representations. The A8 designation refers to the quantization scheme's activation precision, not a promise that every runtime executes native INT8 activation kernels. Actual speed and VRAM usage depend on hardware, loader, offloading, and implementation.

GGUF size estimates

The following are preflight estimates, not independently verified final file sizes, except for Q8_0:

GGUF format	Approximate size
Q8_0	20.866 GiB (observed file size: 22,405,029,216 bytes)
Q6_K	~16.520 GiB
Q5_K_M	~15.805 GiB
Q4_K_M	~15.133 GiB

These files retain a substantial number of tensors at higher precision, so their sizes are larger than estimates based on applying the headline bit width to every parameter.

How the GGUFs were made

The GGUF ladder is produced from the original P-H3-Image BF16 safetensors checkpoint using a standalone, explicit memory-mapped (mmap) conversion pipeline. The converter:

• Maps the source checkpoint read-only rather than loading the entire model into system RAM.
• Processes tensor data in bounded chunks.
• Uses native GGML quantization for Q8_0, Q6_K, Q5_K, and Q4_K tensor blocks.
• Writes GGUF tensor data directly to disk.
• Preserves original tensor names and marks the GGUF architecture as minimax.
• Retains selected non-transformer, normalization, modulation, and other sensitive tensors at F16/F32.

The source checkpoint contains 532 tensors: 420 BF16 and 112 F32. Of these, 200 primary transformer weight tensors are selected for GGUF quantization under the current policy; 220 are stored as F16 and 112 as F32.

The GGUF converter is independent of ComfyUI's node installation. That does not imply that every existing GGUF loader can execute these files.

W6A8 and other safetensors variants

The W6A8 variant was produced using streaming quantization with ConvRot group size 256 and weight quantization group size 32. The reported conversion covered 200 transformer layers, with mean reconstruction relative L2 error 2.344% and maximum 2.411% on the measured quantized tensors. This is a quantization reconstruction metric, not a perceptual image-quality benchmark.

W4A8 and INT8 ConvRot use their respective quantization layouts and require compatible implementations. Do not assume a standard safetensors loader can decode packed quantized weights.

Using these files in ComfyUI

1. Download the P-H3-Image diffusion model variant you want to test from this repository.
2. Place the model in the location required by its compatible quantized-model loader. For diffusion GGUF loaders this is commonly ComfyUI/models/diffusion_models/; follow the specific loader's instructions.
3. Use the P-H3-specific text encoder, not an unrelated MiniMax/H3 text encoder. The P-H3 Comfy distribution provides text_encoders/p-h3-image_text_encoder_w4a8.safetensors.
4. Supply the compatible VAE and the correct P-H3 workflow/model settings.
5. Test a small generation before relying on the format in production.

Compatibility warning: GGUF architecture metadata and a successful GGUF reader check do not prove that a given ComfyUI-GGUF loader recognizes the P-H3 model architecture or its tensor execution path. Likewise, W4A8/W6A8/INT8 safetensors need an appropriate quantized-weight loader. These formats should be treated as experimental until generation is verified.

Choosing a format

• Prioritize fidelity: Start with Q8_0, Q6_K, or INT8 ConvRot, subject to loader support and actual image comparisons.
• Balance size and fidelity: Compare W6A8 with Q5_K_M.
• Minimize model storage: Try W4A8 or Q4_K_M, then check for visible artifacts.

These are starting points, not benchmark rankings. GGUF bit width, activation-aware safetensors quantization, runtime memory consumption, and generation speed are different dimensions; compare them using the same prompts, seeds, resolution, workflow, and hardware.

Testing and limitations

• Q8_0: Converted to a 22,405,029,216-byte GGUF; parsed successfully with the Python gguf reader, which reported 532 tensors and minimax architecture metadata.
• Q6_K / Q5_K_M / Q4_K_M: Preflight native-quantizer checks passed. Final output and image-generation validation were not yet confirmed when this card was drafted.
• W6A8: Streaming quantization completed; reconstruction error was measured as described above. This is not an end-to-end image-generation quality result.
• W4A8 / INT8 ConvRot: Included as available variants; no P-H3 end-to-end benchmark or compatibility claim is made here.

If you encounter a loader error, include the exact filename, loader/node version, ComfyUI version, and traceback when reporting it.

Credits and licensing

• Base model and ComfyUI distribution: PrunaAI/P-H3-Image-Comfy
• Quantizations and packaging: Rebel AI
• Quantized model repository: realrebelai/P-H3-Image_GGUFs

Licensing: Refer to the upstream model's license and usage restrictions. Quantization does not remove or supersede upstream licensing obligations. No new license is asserted for the original model weights by this README.

Follow Rebel AI

• YouTube — @realrebelai
• X — @realrebelai
• Hugging Face — realrebelai

Rebel AI — Making powerful AI models accessible to everyone.