lodestones

/

Chroma

like 791

I'm surely not an expert and still new to Chroma, and would love to read advice from others, too.
But I've recently started exploring other samplers / schedulers, so here are my (very humble and early) feedback and 2 cents:
Note: I'm heavily biased toward "photo rendering", and never generate illustrations, drawings, anime, ...

1. the standard: the most common, as I read in posts, are euler / dpmpp_2m (or euler_a), with beta / sgm_uniform (or others).
   They work pretty well, but they may tend to create drawings, although it has clearly improved with latest versions.
2. I tend to like the ddim_uniform scheduler: a more realistic rendering, and fine details. But it can become quite creative, in a good or bad way...
3. Many samplers will do, with the notable exceptions of the "sde" versions of DPM. But there is a but, see below (point 6)...
4. I very recently explored the ipndm sampler, very interesting imho. You can lower the CFG (thus improving realism), while keeping a good prompt adherence. And it adds some very interesting details, making the scene more lively and immersive.
   But this beast is a bit wild, and I haven't found a "perfect param" yet. But on case-by-case, can definitely bring a nice touch after tuning (probably good for refining / upscaling). My 2 cents about it:

• with beta / sgm_uniform: not very interesting. Probably CFG ~2.0-3.0, but in some cases I could go down to 1.6. But tends to get messy with deformities, bad hands, loss of proper lighting, and so...
• with ddim_uniform: very interesting! CFG a bit higher, around 2.5-3.5. But be prepared: can be very creative. But it may definitely add something I couldn't reach with other combinations.

5. explore the _cfgpp (CFG++) versions of samplers. Not spent a lot of time on those, but they can surely bring. But adjusting the CFG is not so simple.
6. you may check the Skimmed CFG extension. It allows using very high CFG (up to 100...) without burning. And can be a solution to improve prompt adherence, or use the "sde" DPM samplers. I didn't invest a lot of time on it, so cannot properly comment.

Sorry it's long and not the most rigorous, but I hope you'll find some interesting ideas to explore.
If I can in the near future, I will try to analyze with XY Plot grids to get a proper comparison. Please tell me if you would be interested.

any ODE based sampler should work
and for the scheduler sigmoid or beta curve should generate pretty good images for fewer step count