As I don't have very advanced hardware, if you could provide me Buzz to train Lora I would appreciate it. With 500-4000 buzz I can train a new Lora to merge with the model or make it available individually. My PC becomes a nuclear reactor training any type of lora or checkpoint.

Available at TensorArt: https://tensor.art/models/764428935426615833

Stable Diffusion prompt generator AI Tool Bot: https://hf.co/chat/assistant/65c5622d649836a37deb39f6

Default SD 1.5: Use the "DPM++ SDE Karras" or "DPM++ 2M Karras" sampler with only 20 steps for better quality. "DPM++ SDE Karras" has better quality but is slow.

Default SD 1.5: Use the CFG scale between 5 and 9. CFG scale at 5 is recommended. Stable Diffusion default value is 7.

SD 1.5 LCM: Use the "Euler a" or "LCM" sampler with only 8 steps for better quality.

SD 1.5 LCM AND SDXL Lightning: Use the CFG scale between 1 and 2. CFG scale at 2 is recommended.

Pony SDXL: Use the "Euler a" or "DPM++ SDE Karras" sampler with 20-30 steps for better quality.

LCM with DPM++ SDE Karras: Sampling steps 8, Hires. Fix enabled, upscaler latent, Hires steps 2, hires. denoise 0.45, upscale by 1.05, cfg scale 1.5.

LCM with DPM++ 2M Karras: Sampling steps 8, Hires. Fix enabled, upscaler latent, Hires steps 2, hires. denoise 0.55, upscale by 1.05, cfg scale 1.5.

SD 1.5: Resolution between 512 and 768 using "Hires. Fix" or other form of upscaling. The 600x600 resolution is the one that has worked best.

If you use **"**Hires. Fix", use "Upscale by" with the value 1.5 with denoising at 0.6.

Use (word) for larger weight and [word] for smaller weight and word:0.5 to set the weight manually.

Use Token Merging Ratio 0.5 on optimization settings.

If you have a GPU with low VRAM, use --medvram or --lowvram on command line arguments (webui-user.bat on Windows).

Enable fp8 in the latest versions of Web UI to save a lot of ram. It can speed up SDXL a lot depending on how much ram and vram you have.

If you are having a lot of trouble running SDXL in the Web UI. Try using Fooocus

Extensions I recommend: Controlnet, Booru tag autocompletion, CivitAI Browser+,Ultimate SD Upscale,Lobe Theme

Recommended negative prompts to improve human images: amputee, deformed body, long neck, extra fingers, bad body proportions, mutated hands, mutilated, mutation, ugly, fused fingers, malformed limbs, extra heads, disfigured,

I would like to thank the following projects that were important to the community and helped improve Stable Diffusion to achieve greater levels of realism:

Special thanks to other projects from anime and artistic models:

This model is capable of generating very realistic images that can confuse people's critical sense, please do not use it irresponsibly, harming ordinary people or public figures. Also avoid its use in electoral campaigns, promoting or damaging the image of candidates. Although the field of artificial intelligence is poorly regulated, its improper use can cause legal penalties even in countries with little regulation in the area. Be very careful to not commit crimes in your country and not to promote injustices against innocent people. I do not authorize the use of this model for Lora training using the face of real people without authorization from the biological owner. I do not authorize the use of images generated by this model for scam applications and the dissemination of false information. If the image is not included in the cases of unauthorized use but for some reason may cause confusion among the public and may end up being considered true, causing some form of misinformation, place a notice that the image is generated by AI.

I do not allow the upload of any of my models on Tensor Art. I only do not allow it there because the policy of duplicate models on Tensor Art prevents me from uploading again if someone has already done it.

As it is unfeasible to moderate all NSFW posts using real people's Loras. I have enabled to automatically hide all NSFW posts.

This version is the final version after 20 beta versions. This version features more realistic face textures, more consistent scenes and more realistic contrast. Unfortunately it has lost some of the Flux style that it had in version 11, to return to the Flux style it would be necessary to train the model again and this would not necessarily improve the model.