Qwen Image 2.1 Turbo - extend 'magic' 8 steps sigmas further
retained text
So, I've decided to try Qwen Image 2.1 Turbo with the following 'magic' sigmas, taken directly from their Diffusers implementation: `[1.0, 0.978453, 0.954180, 0.926626, 0.895080, 0.845148, 0.704534, 0.414568, 0.0]` This works okay, but for fine details we need slightly more steps (for better skin, better composition etc.) And to my knowledge, there's no working way to interpolate the resulting curve to any N steps (e.g. 10, 12, 14). There is 'Custom Sigmas' RES4LYF node, but it is bugged (as I will show later). Here's the original 'magic' 8-step curve: ['magic' 8-step curve](https://preview.redd.it/2kg2j7angquh1.png?width=340&format=png&auto=webp&s=6189c8fa5f1eb004287efa00de5acb3779521d05) If we use RES4LYF 'Custom Sigmas' node and try to interpolate to, say, 14 steps, we get this: [curve is screwed](https://preview.redd.it/732x6qf0hquh1.png?width=337&format=png&auto=webp&s=b6c913017c81657c52e796652893ec4fa230240b) So, I had to franken-craft a solution, but I did not want to create a custom node just for this purpose. Meet a 'Magic AI Slop Scheduler for Qwen Image 2.1 Turbo'😆 [Much better, huh? Interpolated to 14 steps](https://preview.redd.it/uigkzy4hhquh1.png?width=625&format=png&auto=webp&s=f741bd5e751330e0b385cfd8f5d3d280ab6c8937) Note that curve looks exactly the same now, but it has more steps (14 in this case). For this, you gonna need **RES4LYF** nodes, **KJ** nodes and **rgthree** nodes. The last is the most important one: it has a 'Power Puter' node which allows to make these calculations (but, probably, many 'expression' style nodes from elsewhere would do the same) And indeed, this works for 10 steps, 12 steps, 14 steps (and probably beyond, but there's no sense to run a Turbo model for more steps). So, if you: \- Use this 'scheduler' \- Use CFG > 1 \- 12-14 steps \- Samplers (euler\_ancestral \[the best IMO\], euler, er\_sde) ... you might actually end up with some beautiful and detailed coherent pictures. For example: https://preview.redd.it/pcdf608liquh1.png?width=1248&format=png&auto=webp&s=4283bcc8851ae6052e158577b6aa92931d1155aa https://preview.redd.it/ahrbeebmiquh1.png?width=1248&format=png&auto=webp&s=0c7974c0626e81fa4d5d0b7a0719320bd6602b27 https://preview.redd.it/ywfpp31niquh1.png?width=1248&format=png&auto=webp&s=0d727e4364ebcaff000e97b85e879f775468d1b3 https://preview.redd.it/6dh6xj3oiquh1.png?width=1248&format=png&auto=webp&s=ed71d63ceb359ec556837e7d137a702ad1ae7b4c 1girl, 2 girl is on purpose, maybe more people will actually read this 😆 For a workflow, head to Civitai and download original 2 megapixel gens, just drag those images into ComfyUI. There's a pretty extensive explanation inside on what actually happens and how to use. All pics above generated using 12-14 steps and CFG from 3.5 to 5, @ 2 megapixels. Links: [https://blobs-b2.civitai.com/file/blobs-managed-public/Q054FWT4DMQR88GYN0PXJGQ030](https://blobs-b2.civitai.com/file/blobs-managed-public/Q054FWT4DMQR88GYN0PXJGQ030) [https://blobs-b2.civitai.com/file/blobs-managed-public/V47QEA6MNP2XZAXPFFSGAHXZQ0](https://blobs-b2.civitai.com/file/blobs-managed-public/V47QEA6MNP2XZAXPFFSGAHXZQ0) [https://blobs-b2.civitai.com/file/blobs-managed-public/CSDBTWXEJ1KS54MCZN9XK2S6K0](https://blobs-b2.civitai.com/file/blobs-managed-public/CSDBTWXEJ1KS54MCZN9XK2S6K0) [https://blobs-b2.civitai.com/file/blobs-managed-public/C30VR7Z4ZHRBH8FW5RZ1YSFGC0](https://blobs-b2.civitai.com/file/blobs-managed-public/C30VR7Z4ZHRBH8FW5RZ1YSFGC0) Bonus feature: I've included a 'fake SHIFT' so you can nudge the curve a little bit; negative values = more fine details, like '-1.0' or '-2.0'. But do not overdo it: we don't want to deviate from the original curve too much. Set 'fake SHIFT' to '0.0' to disable it. https://preview.redd.it/pz6g8uw8kquh1.png?width=742&format=png&auto=webp&s=3ace3e3e0eff1ab84fd964f0c45705beb8db7521 Bonus tips: \- please, oh please, **do not use CFG=1**. You'll get underbaked yellowish gens and you're missing out on a negative prompt (which works VERY well) \- always start with less steps (10), only increase (up to 14) if composition is wrong or there's not enough fine details \- then adjust CFG, 3.5 - 5.0 works just fine, but this is very specific for each particular gen \- if there's 'too much details' (e.g. skin is too detailed or grainy) - reduce steps first, then increase 'fake shift' (set it to '1.0' or '2.0' etc.) \- vise versa: if skin is too plastic, increase steps, then decrease 'fake shift' (set it to '-1.0', '-2.0' etc.) \- most important: Qwen Image 2.1 Turbo has to be prompted PROPERLY. If some of you remember original Chroma prompting, you probably know what I mean - using exact phrasing, being specific, avoiding slop tags like '1girl, masterpiece' etc. - are all the keys. No gen params will ever fix bad prompting. Have a nice day! EDIT: for those folks who prefer cleaner output without excessive noisy details, here's somewhat Krea-like output (just less steps, and 'fake shift' at 2). Hey, it even generates faster 😄 https://preview.redd.it/57mjc7mp8ruh1.png?width=1248&format=png&auto=webp&s=a841a713cf3ca77f4bb767814823159574b1c7bb
document fetched 2026-10-12 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1x2v049/qwen_image_21_turbo_extend_magic_8_steps_sigmas