Awful Image Generation

[shaun-ba]

I mean what is this? I've followed all guides and reported various bugs for the past few days and this is the outcome?

[lawrence-cj]

Try something without anatomic? We are trying to improve human body and human face in our next release.
Other kinds of images are ok I think. Refer to: https://nvlabs.github.io/Sana/

[shaun-ba]

@lawrence-cj I cannot get anything to look realistic not just humans. See below for prompt "large oak tree". I thought your goals were to beat Flux for realism with 4x speed, but there isn't a single prompt that looks realistic to me, so I must be missing something?

[shaun-ba]

Here is one from your homepage, prompt "a cyberpunk cat with a neon sign that says "SANA""

[shaun-ba]

@lawrence-cj I've noticed that lowering CFG to 2.0-2.5 makes a massive difference with realism, under 2 then images don't look right, over 3 and they are as above kind of over exposed with lots of highlights on the edges of the detail etc.

Below is CFG 2.5, but even on this can you see on the ball of yarn the highlights on the top are blown out?

[lawrence-cj]

Oh, if you want the content to be more realistic, then the prompt needs to be refined.
Here is a simple way to rewrite your prompt:

Sana/app/app_sana.py

Line 64 in bcad148

"prompt": "cinematic photo {prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed",

BTW, we can't guarantee that Sana's smaller models will beat the FLUX just yet. Sana's larger model sizes and better VAEs are also in development and we have got pretty cool results. What we can guarantee is that we will keep improving and keep everything efficient. Maybe we still need time.

At last, thank you for your time to play with Sana. We will keep working hard.

[lawrence-cj]

Besides, rewriting the prompt to a longer one will make the quality better in our experiment.

[shaun-ba]

@lawrence-cj I am used to writing very complex prompts with Flux with layering, background, middle ground, foreground etc. I was just testing as even in flux with a few words you get very good realistic results still. Using longer prompts there is generally more beneficial for finer control of the environment, subject, focus and such.

Using a few of your keywords above I don't get any better results. CFG does hugely effect the image though, I don't know if this is expected behavior or something odd going on with Comfy or schedulers etc?

Don't worry, I know this is a WIP and the speed is honestly outstanding, hopefully you can get "women lying on grass" on par with Flux!

CFG 4

CFG 2.5

[lawrence-cj]

Using a few of your keywords above I don't get any better results. CFG does hugely affect the image though, I don't know if this is expected behavior or if something odd going on with Comfy or schedulers etc?

@shaun-ba Thanks for your valuable testing results. I think "CFG will affect a lot" is a normal phenomenon. Higher CFG will make the image saturated and lower will make the image style more diverse and unstable. Is there no big difference between CFGs in FLUX? What kind of CFG is better during your testing?

Besides, our official Flow-DPM-Solver scheduler is not supported in ComfyUI for now. For the KSampler-Euler in ComfyUI, I didn't do a lot of experiments, I just made sure the workflow was working.

[bghira]

@lawrence-cj as flux is distilled, it has no CFG

[lawrence-cj]

@lawrence-cj as flux is distilled, it has no CFG

Emmm, Flux-schnel has it.

[bghira]

that is fake guidance, not real cfg. it is distilled into that input argument and it behaves nothing like true cfg; you can use it up to 20.0 and beyond without ruining the image