# Qwen-Image GGUF in ComfyUI

Session context: user asked whether to run Qwen image generation locally in ComfyUI, and whether it can run in parallel with a local video model.

## Practical model choice

For local, free/open-weight image generation in ComfyUI with GGUF, prefer the newer **Qwen-Image-2512-GGUF** over older Qwen-Image variants when the goal is general image generation with strong realism, prompt following, and text-in-image ability.

Useful family split:

- **Qwen-Image-2512-GGUF** — primary txt2img choice; Apache-2.0; practical GGUF path for ComfyUI-GGUF.
- **Qwen-Image-Edit-2511-GGUF** — use for edit/reference-image workflows and changing text/details inside existing images.
- **Z-Image Turbo GGUF** — lighter/faster fallback for low VRAM or quick experiments; lower ceiling than Qwen-Image-2512.
- **FLUX.2-dev-GGUF** — quality alternative but heavier and less convenient/licence-practical for this pipeline.

## VRAM/quantization guidance

Start conservative and ramp only after a small smoke test completes:

- **8–12 GB VRAM:** Q4, low-VRAM/offload modes; expect slow runs.
- **16 GB VRAM:** Q5 or Q6 tests if the rest of the workflow fits.
- **24 GB VRAM:** Q6 or Q8 is a sensible quality/stability balance.
- **<8 GB VRAM:** prefer a lighter model such as Z-Image Turbo GGUF.

The model is a large image-diffusion stack, not a normal LLM. The GGUF file size is not the full runtime memory story: text encoder, VAE, resolution, batch size, previews, and any simultaneously loaded video model can push total VRAM/RAM much higher.

Approximate Qwen-Image-2512 GGUF file sizes seen in practice:

- Q4: ~11.5 GB
- Q6: ~16.8 GB
- Q8: ~21.7 GB

## Parallel with video models

Do **not** assume Qwen-Image and Wan/Wan2.2 video can run concurrently on a 16 GB class GPU. Treat them as separate heavy workloads unless proven otherwise. When answering this class of question, be explicit: **yes, run Qwen-Image locally if it improves keyframes; no, do not plan on true simultaneous GPU residency with Wan video on 16 GB hardware. Queue/orchestrate the jobs instead.**

Recommended orchestration:

1. Run image/keyframe generation as a separate ComfyUI job/profile.
2. Unload/free models before video generation:
   `POST /free {"unload_models": true, "free_memory": true}`.
3. Then run Wan/Wan2.2 I2V/TI2V.
4. Only test true parallelism after single-job smoke tests are stable, and monitor both VRAM and system RAM.
5. If true parallelism is required, prefer separate GPUs/hosts or separate ComfyUI instances pinned to different devices. CPU/offload parallelism is usually a throughput trap for this user’s quality-first video pipeline.

For the user's Retention Studio preference, this fits the quality-first pattern: generate/curate keyframes with Qwen-Image, then feed approved frames into Wan2.2 I2V rather than relying on raw T2V.

## Setup notes

- Install **ComfyUI-GGUF** custom node support before loading GGUF checkpoints.
- Verify exact folder expectations from the node's loader dropdowns via `/object_info`; do not guess model subdirectories.
- Use `hf download` rather than deprecated `huggingface-cli download` in new docs/scripts.
- Run a tiny smoke workflow first: low resolution, 1 image, low steps, fixed seed, then ramp resolution/quantization.