Prompt-Enhancers-INT8-ConvRot-Qwen-Image-2.1

INT8/ConvRot prompt enhancers — derived from Qwen Image 2.1

Built with Qwen. Community conversions, not an official Qwen release and not endorsed by Alibaba / Qwen. Distributed under the included Qwen Research License for non-commercial use; see LICENSE and Notice.

These are prompt-enhancement models, not diffusion checkpoints or image-conditioning text encoders. T2I expands a text prompt; I2I expands an editing instruction using reference images. They return structured text intended to be parsed into a rewritten prompt and aspect-ratio advice.

The source authors describe their derivatives as Heretic/abliterated. This conversion does not independently establish those behavioral claims. The I2I source is the named darrellbest derivative, not a conversion of pottokao's I2I GGUF.

Conversion

Modified files: both model.int8_convrot.safetensors files. BF16 source tensors were converted using Comfy-Org/comfy-model-tools, quant_int8_convrot.py, with the output head explicitly included (--include '^lm_head$') and default non-MSE-clipped quantization. The token embedding is also quantized. An external reader supplies the original indexed shards, and oversized matrices are processed in row batches using the upstream math to fit a 16 GB GPU. No model retraining or additional ablation was performed.

Each checkpoint contains 310 quantized weight bundles. Small/non-matrix tensors retain their original precision. Each is approximately 9.47 GB. conversion.json records immutable upstream revisions, converter revision, output SHA-256 checksums and tensor-payload checksums.

Usage and support files

Use a runtime with native Qwen Image 2.1 prompt-enhancer architecture support and Comfy INT8/ConvRot SafeTensors loading. These quantized files are not ordinary Transformers BF16 shards or GGUF files. In ComfyUI, use the matching native prompt-enhancement workflow/model slot, not the image generation text-encoder slot.

Each task folder includes its source tokenizer, processor, configuration, chat template and required system_prompt.txt. Keep the T2I and I2I instructions paired with the corresponding model. Do not interpret the intermediate reasoning text as the final rewritten prompt.

Validation limits

Both converted checkpoints loaded and executed in Comfy's native implementation and an independent native implementation. For the selected T2I and one-reference I2I cases, captured input tokens, embeddings, prefill hidden states and nine output-score vectors matched exactly; I2I vision features matched exactly too. Each comparison exercised 128 teacher-forced decoding steps. Independent sampled runs returned valid final answer JSON with rewritten prompts and aspect advice.

This establishes compatibility on those cases. It is not a broad quality evaluation, a refusal-behavior evaluation, a BF16-versus-INT8 generation study, or a claim of identical speed/memory on other hardware. The source license and intended-use restrictions still apply.