← sources

Qwen Image 2.1 + Turbo Lora - A curious case (settings)

source: reddit (reddit:StableDiffusion)published 2026-09-26 01:07ingested 2026-09-27 02:38

retained text

EDIT: THIS is totally unscientific hack! works only with my particular inference settings with deis\_2m (RES4LYF bongmath variant + beta\_57 or beta 0.7/0.85) with lora strength around 0.5, no custom sigmas and no special lora load nodes AND with comfy kitchen attention. Discovered by a pure accident. This is shared ONLY as a curios peculiarity, as this might break other things like using extra loras etc., but when used as-is this provides the BEST results considering quality/steps ratio. I encourage you to get embedded workflows from links to full images below and experiment for yourself (simplified workflows on purpose just for sharing sake).

EDIT 2: Apparently what probably might be happening here is just when you append an empty cond you are padding your main cond with empty stuff,  so you just dilute your main prompt, so attention mechanism just smoothes out the final result. Something like that...

Many of you are enjoying this particular Turbo Lora: (EDIT: specified wrong version initially) [https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/blob/main/Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors](https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/blob/main/Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors) (or r128 variant)

I noticed a strange and amazing thing with my particular inference settings: when I use this lora at 14 inference steps (strength 0.5) - 'by default' - it produces fried results, like this one:

[Fried, 14 steps with lora, 'default'](https://preview.redd.it/cje6wuesjrrh1.png?width=1184&format=png&auto=webp&s=18cf843467b2e262d67e2b51b52cc73e5389ee7e)

But! When I concatenate an extra empty conditioning to main prompt (first positive prompt, then concat an empty conditioning), it produces very nice results every time, like this:

[Nice, 14 steps with lora + empty extra cond](https://preview.redd.it/scp2ys11krrh1.png?width=1184&format=png&auto=webp&s=65135588fca8ffe75c1d3498af8735a54bffa509)

For reference, here's a gold-standard 50 step variant with no lora and without an extra conditioning:

[50 steps, no lora, no extra cond](https://preview.redd.it/7xrrbk26krrh1.png?width=1184&format=png&auto=webp&s=7bca9a307f242ee30a07666856dfcd4ceb3604da)

Full glory variants to see for yourself, all with workflow included:  
[https://blobs-b2.civitai.com/file/blobs-managed-public/2FWVY80H95GHY72RT15BM2JX40](https://blobs-b2.civitai.com/file/blobs-managed-public/2FWVY80H95GHY72RT15BM2JX40)  
[https://blobs-b2.civitai.com/file/blobs-managed-public/1NAM6MC808N6D14ZG7B5CQY910](https://blobs-b2.civitai.com/file/blobs-managed-public/1NAM6MC808N6D14ZG7B5CQY910)  
[https://blobs-b2.civitai.com/file/blobs-managed-public/FAT7A46JS56MPSG8GSH72GFD60](https://blobs-b2.civitai.com/file/blobs-managed-public/FAT7A46JS56MPSG8GSH72GFD60)

All you have to do is to enable/disable these two purple nodes:

[Nodes to toggle](https://preview.redd.it/0z2ltmb3lrrh1.png?width=933&format=png&auto=webp&s=fa29f64113c14293d131fe96b1797c4d149d15d3)

I do not know why it behaves like so, but it does. And I believe that 14-step Lora version with an extra empty cond is almost on par with full 50-step version. 🤷‍♂️

You can find all setting in the workflows embedded into 3 linked images above.

document fetched 2026-09-27 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wqdazi/qwen_image_21_turbo_lora_a_curious_case_settings