Sign In

Real Dream

1.6k
40k
853k
687
Verified:
SafeTensor
What did you think of this resource?
Type
Checkpoint Merge
Stats
2,894
Reviews
Published
May 1, 2024
Base Model
SDXL Lightning
Hash
AutoV2
8DDC302D1F
default creator card background decoration
sinatra's Avatar
sinatra

As I don't have very advanced hardware, if you could provide me Buzz to train Lora I would appreciate it. With 500 buzz I can train a new Lora to merge with the model or make it available individually.

Available at TensorArt: tensor.art/models/680175287830377266

Real Dream Hugging Face demos:

tinyurl.com/realdreamlcm1demo

tinyurl.com/realdream10demo

tinyurl.com/realdream8demo

You can find old versions at Hugging Face: tinyurl.com/hf-realdream

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

Take a look at my bot that solves math problems using Python: https://hf.co/chat/assistant/663652661ed733523d962772

Recomendations:


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.

Tips for a better prompt:

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

Recomendations for improvement on performance:

Use Token Merging Ratio 0.5 on optimization settings.

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

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:

  • Realistic Vision

  • epiCRealism

  • epiCPhotoGasm

  • QGO

  • Analog Madness

  • majicMIX realistic

  • CyberRealistic

  • Juggernaut

  • LCM Lora

  • SDXL Lightning

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.

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