Real Dream Website (Faster download)
Available at TensorArt: https://tensor.art/models/899250342552672189/
Comfyui Workflows:
Anima text2image with Loras Workflow
Anima img2img (denoise) with Loras Workflow
Klein 9B Edit with Loras Workflow
Klein 9B img2img (denoise) with Loras Workflow
Z-image-turbo img2img workflow
Z-image text2img with loras workflow
Z-image De-Turbo with loras workflow
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. Do not create images that can embarrass and damage the public image of people. This can cause serious psychological harm, including deep depression. I do not authorize the use of images generated by this model for scam applications and the dissemination of false information.
During the fine-tuning of AI models, I only use synthetic images that do not infringe any copyright, personal image rights, or laws. The models are merges of open models from different sources and may contain diluted data remnants in the neural network. I am not responsible for possible data remnants since they are diluted, and it is not humanly possible to track the existence and origin of this diluted data.
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.
I have enabled to automatically hide all NSFW posts for several reasons that are not worth listing. I hope you have patience and understand.
Description
FAQ
Comments (23)
Don't you check your DM's
Unfortunately, I haven't been responding to messages because I have a lot and I haven't had the focus to respond.
@sinatra I have sent you 13k credits seeking help on something so please can you contact me on [email protected]
20 is the best one yet! Awesome job!
Will the "flux chin" phenomena be as prevalent as the preview images?
@sinatra Real Dream is awesome but Real Dream with Qwen Image should be even better. For speed between Flux and Sdxl and with better quality then both. I wish you think about version for Qwen. Cheers.
Not sure what I'm doing wrong in comfy, using the workflow embedded in your example images (and other uploaded images), I always generate a black image when using RD8 SDXL (not pony). I have no problems generating images using other models.
Same, until i've changed clip skip to -2 . Struggled for a lot to figure this out.
@johnnys_ai Yea, some models have required clip skip, not optional.
@johnnys_ai thank you that was the issue
Does it not have a recommended positive prompt? Only a negative one
do you have a fp16 or 32 version somewhere. you have great models and fp8 is selling yourself short
You should actively compare a fp8 vs fp16. Most of the time there is no difference at all because of a bfloat16 conversion
@Synop you sent me down a rabbit hole on this one! it ends up that unless you are doing extremely high end or scientific research fp8 even holds up to fp32 if tuned correctly 🤯AI FOR THE WIN!!!
@Cosmic_Crafter its even worst than that, for most of us, there is no need to tuned it, comfyui have some technologies than make everything converted automatically. I guess it depend of material. https://www.reddit.com/r/comfyui/comments/1nsq3gc/exact_same_result_bug_with_fp32_an_fp8_models/
@Synop yeah comfyui has model managing and OOM protection built in so it might be doing all the block swapping for us. I want if the encode and decode process is the quality bottleneck if the encoding process smashes everything down to 8 bit then unpacks it is it lossless? does the color profile compression hurt the final image?
@Cosmic_Crafter @Cosmic_Crafter These are genuine questions that I ask myself as well. Unfortunately, due to a lack of time and knowledge, I don't have the answers. In any case, it's interesting because resolving and understanding these elements could allow us to develop better quality creative processes in the future. If you find any answers, please don't hesitate to mention them; I'll do the same.
@Synop Comfy's dev team recently downgraded LORA processing to 16-bit. It used to be fp32. Unsurprisingly, there doesn't seem to have been a setting added in the GUI to switch it back. FP8, though, is quite a bit worse than FP16. If people want small formats they should stick with Q8.
@Synop That hasn't been my experience at all. Every set of examples of I have seen shows FP8 changes the output a lot. Q8 has been the closest to FP16 in every test I have seen.
@Cosmic_Crafter I'm not sure what the question is but all of the quantizations are lossy. The key is using one that doesn't significantly affect the quality. Most seem to agree than FP16 and BF16 are sufficient for most processing and Q8 is okay if you need to go lower. However, back when I was using SD 1.5 I achieved clearly better results with FP32.
@ss9999 if you dont have the same experience its because you dont have the same setup. This mean your hardware make your comfy act be receptive to differents quantization. Mine isnt. Its the same for everything
Hello, what are your tips for avoiding vertical and horizontal streaks appearing with the Flux v2 model? Thank you.
Edit: Find solution: https://github.com/tritant/ComfyUI_Remove_Banding_Artifacts
Hey! Might be a bit late, but if you’re using ComfyUI, the "DyPE for Flux" node solves that problem. It does an extrapolation during hires fix and gets rid of those vertical/horizontal streaks.
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.