A word of encouragement

[mr-username-iam]

I've been browsing this repository, trying to understand the current state of the beta and the questions, suggestions or issues people might be having, before installing myself, and just wanted to say:

Kudos to the developers behind this project. Not only is it a fantastic idea, with great potential going forward, and one that deserves more attention, but also one that is actively improved upon in significant ways; model conversion, one or two step generation, real time text to image, etc., these features and therefore software stand out as a cut above the rest, even some of the better known names which have been around longer.

I appreciate how you've prioritized efficiency and adoptability, even if at present it is a little intimidating for less computer savvy folk like myself.

[rupeshs]

Thank you so much for this wonderful message,it truly means a lot!
I created FastSDCPU with the goal of making image generation accessible to everyone locally, not just those with high-end GPUs, so hearing that the project's direction that just resonates with you is incredibly rewarding.

[mr-username-iam]

Glad to hear it. I really feel this area is underexplored and generally overlooked. The projection is very solid, the market massive.

GPUs are and will remain "premium" items, but CPU compute power increases every year, and it won't be long until local inference is widely accessible, if we consider today's models to be acceptable. Everyone with a laptop or cellphone will be able to run local models, if the software is also accessible.

[ruberkel]

It's fantastic. I am amazed at Fastsd cpu. recently I had to buy a cheap thinkpad, So i changed a 4gb 3050 Nvidia for a UHD integrated ... because the old laptop died. Now, I have an intel core I5-8365U 1.60 ghz and an integrated UHD 620. I'm no programmer, but Im good with prompts.
I managed to get perfect images even on sd turbo with positive (neg) prompts with words like avoid or without... etc. Using the upscaler, as It really acts as an Adetailer with simple prompts to fix hands, or whatever and works well and fast. Marvellous! I get below 3/its!

Very grateful to @rupeshs and collaborators. thanks!

PD. I will post my link to my sd turbo images for you to see. you can ask me how to improve on it.

[rupeshs]

@ruberkel Thank you so much for the kind words! 😊
It’s fantastic to hear that FastSDCPU is working well even on an i5-8365U with Intel UHD 620. Making image generation usable on ordinary and older hardware is one of the main goals of the project, so feedback like this is really encouraging.
Please do share your SD Turbo images when you can. I’d be very interested to see the results and hear more about your prompts/settings(or any posts). Your feedback could also help improve FastSDCPU further.

[ruberkel]

Hey, If I can help...

My specs: thinkpad x390, windows 11, 16 gb Ram, i5-8365U with Intel UHD 620.

I use Fastsdcpu always with Gpu on openvino mode, because Cpu is about 2/3x slower.
With 4 steps, 1 cfg, openvino+Taesd, 512px image size, I get 2s/it with SD Turbo, and 5s/it with SD XL Turbo (after first pass/compilation fase). The other models, don't work too well in openvino GPU mode. SD XL Turbo generates a bit low res images so always have to upscale/reimage. The lowres of SDXL Turbo is common knowledge as it was trained at 512px so it looses a lot of detail. Nothing that an upscaling cannot solve. Also increasing CFG makes it slower with any model.

edit. no more black screen with sd turbo XL, works perfect now. probably a mistake on my side. The only one which generates a gray square on gpu mode is SANA model. On Cpu it works well.

Thanks! I will put images when I can, Im still messing with all things, learning.
Regards.

[ruberkel]

Would it be a good idea to make a post titled "post your prompt images here" so ppl can post their experiments on different configs and models.

I will post two images here so you can see. All positive prompts with avoid or without, works, its not placebo, even if sd turbo models do not support negatives, it seems to work most of the time, Lol.

All generations using openvino + Tae, GPU mode (set device=gpu on start.bat), 1 cfg.

model SDXL turbo, I think 10 steps. orig. size 512 upscaled to later 1024.
prompt :"real photo, 1girl, adult woman, upper body, sitting on curb, urban street at night, relaxed posture, leather jacket worn open, (avoid disfigured, avoid hands, avoid mutant, avoid polidactyly)++"

Still not perfect, but you can see the perfect position.

model SD Turbo, steps 12. 512 original, i2i with 0.5 denoise strenght.
Prompt: (close-up), masterpiece, professional photo, ((1girl)), a modern young adult woman, seated on a curb, (avoid separate limbs, avoid missing digit, avoid disfigured, avoid hands, avoid distorted, avoid deformed, avoid mutant, avoid polidactyly)

Some observations. eyes a bit pixelly can be fixed with higher res or upscale, as 512 sometimes is not enough for that detail. Context help to keep less deformities, like in the prompts I just posted, in a curb... etc. I2I or upscaling (sd) recognizes hands and feet even with a cheap overpaint.

The best model overall (for me) is SD XL turbo, and turbo is not bad.
All done in fastsd cpu! (using gpu) in a few minutes of work.

hope you like it. Cheers. should I post more?

[ruberkel]

Another difficult one, usually models as good as SD3 can't do it easily.

Done with only sd turbo, 4 steps, 512px, 1 cfg. openvino+taesd, not bad. You get good "laying" top view poses in maybe 1/10 tries, so I changed wording a little, could be sleeping, too. But those are the limitations of SD turbo. If you are dedicated you can always use controlnet. As always, the placebo positive negatives xd.

Upscaled to 1024 with i2i... not perfect, (I2i insist on some weirdness on the start of the legs, original 512 is perfect. good for a fast upscale. Cheers!

prompt: (top view), ((1girl)), ((avoid multiple or extra limbs, avoid hands or feet, avoid arms or legs)), HDR, UHD, 64K, professional photography, high detail, studio lighting, masterpiece, a woman resting

[rupeshs]

I would recommend to use FLUX.2-klein-4B(OpenVINO),steps 2 to3