Qwen-Image-2.1 on RTX 5080 — Independent Community Benchmark

Hardware

• GPU: NVIDIA RTX 5080 16 GB GDDR7 (Blackwell)
• CPU: AMD Ryzen 7 9800X3D
• RAM: 32 GB DDR5
• OS: Ubuntu/Linux
• Runtime: local ComfyUI workflow

Important disclaimer

These are independent real-world measurements from one RTX 5080 setup. They are not universal benchmarks for Qwen-Image-2.1. Reproduction requires equivalent model files, workflow, software versions, quantization, encoder, VAE, sampler, steps, CFG, resolution and hardware.

Main finding

In this specific setup, Turbo8 was the only tested acceleration route that simultaneously delivered very good speed and exceptional image quality.

The important result is not simply “Turbo8 is best”. The result is:

the acceleration method, CFG, VAE, quantization and encoder combination materially changes the outcome.

Measured configurations

Acceleration methods tested and rejected

Lightning

Rejected in this RTX 5080 setup.

• NVFP4: visible degradation / grid artifacts
• INT8: visible degradation / grid artifacts

Viggle

Rejected after showing the same class of visual degradation/artifacts in testing.

SageAttention

Rejected.

• Approximate measured speed gain: ~7%
• Image quality: grid/artifact problem observed
• Conclusion: the measured speed gain was not worth the visual degradation.

CFG finding

For the tested Turbo8 workflows:

Important difference from the public Turbo8 model card

The public Turbo8 release recommends 8 steps and CFG 1. The RTX 5080 results reported here are empirical results from a different local combination and should not be interpreted as a correction to the official model card.

Public Turbo8:https://huggingface.co/chriswritescode/Turbo8-LoRA-Qwen-Image-2.1

Resolution finding

• 2048×1152: not viable in this setup
• 1344×768: practical upper limit found
• 2048×1152 img2img/refine: attention/offload caused severe slowdown and the process amplified the existing texture/grid artifact instead of fixing it

VAE finding

texture_fix_vae_for_qwen_image_2.1_bf16.safetensors from madebyollin was preferred over the original VAE in visual testing because it produced cleaner fine texture/checkerboard behaviour.

Project:https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1

The VAE is an unofficial decoder finetune. Its project page reports improved rFID versus the original decoder while reconstruction metrics change slightly.

Why these results matter

For Qwen-Image-2.1 on 16 GB Blackwell, optimizing only the diffusion model is insufficient.

The correct benchmark unit is the complete pipeline:

DiT/checkpoint
+ text encoder
+ VAE
+ LoRA
+ sampler/scheduler
+ steps
+ CFG
+ resolution
+ attention implementation
+ cache
+ offload
+ runtime

A method that reduces latency but introduces systematic visual artifacts is not an optimization for production.

External reference: Turbo8

The public Turbo8 project reports 8-step generation and publishes external measurements on a 96 GB RTX PRO 6000 Blackwell. Its reported 1024² ComfyUI result is 2.8 s, but that is a different GPU and configuration.

Article:https://cstech.dev/articles/turbo8-distilling-qwen-image-2-1-to-8-steps

Model:https://huggingface.co/chriswritescode/Turbo8-LoRA-Qwen-Image-2.1

Suggested reproduction protocol

1. Fix one prompt set.
2. Fix seeds.
3. Fix resolution.
4. Change only one variable at a time.
5. Record end-to-end wall time.
6. Record peak VRAM.
7. Save the exact workflow JSON.
8. Compare the same image crops at 200–400% for texture/grid artefacts.
9. Evaluate text rendering separately from photorealism.
10. Publish hardware + software + workflow, not only a screenshot.