This is a simplified guide to an AI model called Qwen-Image-2.1-Turbo-Quantized maintained by Abenzerps. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.
Qwen-Image-2.1-Turbo-Quantized is a set of quantized weights for local text-to-image generation, based on Qwen-Image-2.1-Turbo. It is maintained by abenzerps.
Generate images from text prompts in ComfyUI, with stable-diffusion.cpp, or through MFLUX using the provided GGUF, Safetensors, or MLX weights. The documented baseline is 1024 × 1024, batch size 1, Euler sampling, CFG 1.0, and eight steps.
The model card states that the base Turbo model has no built-in safety checker or content filter. Higher resolutions require more memory and time, and model file size does not represent total VRAM use. The README gives conflicting GPU generation-time estimates (about 10–20 seconds versus about 1020 seconds), so it does not establish a reliable runtime estimate. No measured quality comparison for these quantizations is supplied.
The related Qwen-Image-2.1-Turbo is the upstream checkpoint this repository quantizes. The Qwen-Image-2.1-GGUF is described as quantized from the non-Turbo Qwen-Image-2.1 base, rather than the Turbo checkpoint. Other supplied alternatives include Qwen-Image-2.1-Uncensored-Quantized and Qwen-Image-2.1-Turbo-GGUF; the sources do not provide a controlled comparison of their quality or speed.
Based on a 7B visual-generation architecture. Turbo generation uses eight denoising steps and upstream default CFG 1. Available diffusion-weight formats and listed file sizes: FP8 Safetensors 7.13 GB; INT8 ConvRot Safetensors 7.26 GB; NVFP4 Safetensors 4.20 GB; MLX 4-bit 4.00 GB, 6-bit 5.78 GB, and 8-bit 7.56 GB; GGUF Q8_0 7.59 GB, Q6_K 5.88 GB, Q5_K_M 4.94 GB, Q4_K_M 4.05 GB, and Q4_0 4.05 GB. Companion ComfyUI files include BF16 or INT8 ConvRot Qwen3-VL 8B text encoders and a BF16 VAE. The recommended Turbo sigma schedule is 1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0; nine values define eight steps. No training details or controlled performance measurements are provided.
Provide a text prompt. The documented baseline uses 1024 × 1024 resolution, batch size 1, Euler sampling, CFG 1.0, eight steps, and the Turbo sigma schedule. At CFG 1, negative prompts have no effect. A fixed seed is recommended when comparing outputs.
Generated images; the command-line example saves a PNG file. The README demonstrates 1024 × 1024 generation and suggests trying 2048 × 2048 for more detail, with increased memory and time requirements.
For stable-diffusion.cpp, place one diffusion model, the INT8 ConvRot text encoder, and the BF16 VAE in the runtime's models/ directory, then run:
./build/bin/sd-cli \
--diffusion-model models/qwen-image-2.1-turbo-Q4_K_M.gguf \
--llm models/qwen3vl_8b_int8_convrot.safetensors \
--vae models/qwen_image_2.1_vae_bf16.safetensors \
--sigmas qwen-image-2.1-turbo \
--steps 8 --cfg-scale 1.0 --sampling-method euler \
--width 1024 --height 1024 --seed 42 \
--offload-to-cpu \
-p "A natural light portrait photograph" -o output.png
This example assumes the runtime is built and the three files are in the stated locations.
Q: Which format should I start with?
A: The README recommends Q4_K_M for GGUF or INT8 ConvRot for ComfyUI Safetensors. These are setup recommendations, not claims of superior quality or speed.
Q: Does it need a special Turbo LoRA?
A: No. The README says the full Turbo checkpoint needs no additional Turbo or Lightning LoRA.
Q: Can I use a regular eight-step sampler schedule?
A: The README says a regular KSampler with the simple schedule does not reproduce the Turbo schedule. Use the specified sigma schedule.
Q: Are negative prompts useful at the recommended CFG?
A: No. At CFG 1, negative prompts have no effect.
Q: Can I run it at 2048 × 2048?
A: The README suggests trying that resolution after initial setup, but warns it requires more memory and time. Reduce resolution or use a smaller GGUF if memory is insufficient.
Q: Which runtimes are documented?
A: The README documents ComfyUI, stable-diffusion.cpp, and MFLUX. MFLUX is a separate workflow for the MLX 4-bit, 6-bit, and 8-bit weights.
Click here to read the full guide to Qwen-Image-2.1-Turbo-Quantized