I'm doing my best. It's okay if you don't like this model. I'm always doing my best! Just because I'm a beginner, I won't blindly shrink back.Tested at 4090 and the resolution was 1024*1536. The cover image was intentionally set to 768*1366.
Description
I tried it with the intention of starting again from scratch.
(The hands may be a bit random and the drawing may be a bit bad) (I made it again from scratch)
FAQ
Comments (37)
You're doing great, keep going. Supporting your work
thank you I'll do my best!
it's maybe some wrong.use the same node, but v4 it's not look like v4,seem like v3 and v1f1,but not v3 too.maybe for diffent seed form gpu or cpu or webui a1111 or comfyui? i dont know ,it that generate at civitai web online?
This is only a locally executed image. Take the image and load it into comfy. You will see it.
(If you have any inconvenience, we apologize for not being able to explain it properly. I ran it locally. If you want, I can upload a screenshot or video. I won't lie. I'm a beginner and I won't lie about my choices.)
I was late in checking because I was working on lora for a while. I'm sorry if there was any inconvenience. This version was remade from scratch, so I plan to upload the fix to lora as well.
In general, it is not possible to get the exact same image as the one you see on civitai or from the online generator because of differences in software and hardware.
We don't know if civitai's generator is using GPU or CPU RGN. Try both on your local setup with the same seed and see which is closer.
📂💾🗃️🖼️🎨️📐✨
AssertionError: You do not have CLIP state dict!
same
You have not dealt with the topic and write a rudimentary error message, true to the motto ‘something is broken, do something’. With this information nearly no one can help you, only suggests where the error could be.
That sounds harsh and rude, please don't get me wrong, it's not about stopping you from asking questions, it's about showing you that there is a lack of information here to help you get rid of your error message.
What programm GUI are you working with? SwarmUI, ComfyUI, ForgeUI and so on
What have you already done to solve the error message?
Have you ever used Dr Google to search for a solution? As example, I got a solution to this error message after 5 minutes. In my case, for example, I had forgotten to load the clip_l.safetensors and the t5xxl_fp8_e4m3fn.safetensors in ComfyUI or ForgeUI, or the t5xxl_fp16.safetensors if you want to work with 16FP.
you can find the mentioned files here on CIVITAI or on Huggingface:
https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main
have fun, and let's see which images you have created.
For me, all modells are working
Great work, this is the way to achieve a full flux anime version. SD3 and Flux were not properly trained with enough AI data from Midjourney, Niji, Leonardo AI, DALL-E Image Creator, etc. It's not about the analog anime screencap/digital works data or the versatility with which other XL and pony anime models were trained and work. Rather, it's about creating a solid, adaptable base for Lora styles based on existing AI content. This approach creates a new level of originality with its own rendering principles while preserving most of the AI anime generators' style and behavior.
thank you
can we have a LORA version of this for FLUX?
Since it doesn't contain one or two self-made rollers, please receive the fix rollers I upload. I will upload them little by little.
What is a good GUI to run this model in?
I recommend comfy ui
@lyh12169198 can you share your workflow? it's very helpful to me
@ifinetooqqcom671 If you download the png image above and place it in the workflow, a workflow will appear. It is a basic workflow.
This might be the best Flux model of artistic, semi-anime style. Reminiscent of MidJourney Niji. Great work and thank you!
Amazing Model 10/10 ♥
great work. are all trigger words needed? are they different?
You can use the same trigger word as before, but you don’t have to use it.
This is the first time I post a comment about a checkpoint or a lora. But I find your work so remarkable that I can't help but congratulate you for your work. Sincere congratulations, this checkpoint opens new and quality dimensions.
Thank you so much. I'm a beginner, but I'll do my best
Thanks a lot, man, that's the only anime flux model which is really usable and the results are very good. It also works great with NSFW MASTER FLUX lora v1. With a power as low as 0.5 you mostly maintain the style and you get uncensored characters
Some great images can be generated using Bible verse prompts, like the ones below. Give it a try!
https://www.openbible.info/labs/ai-bible-art/?filter=
I cant make it work. I have the recommended VAE installed and seted in Forge, but it doesnt work
Have you set up "FLUX dev in forge with encoders and VAE" under "Installation of Models" on the following site? This setting is required to run the standard Flux models.
https://sandner.art/flux1-in-forge-ui-setup-guide-with-sdsdxl-tips/
@Urahotom Thank you so much for the post!
After many reinstalls of the 6 blue screens of death, I was able to get FLUX to work!!
I had a mistake:
"AssertionError: You do not have CLIP state dict!"
Installing VAE helped me:
ae.safetensors
And text_encoder:
clip_l.safetensors
gitattributes
t5xxl_fp8_e4m3fn.safetensors
t5xxl_fp16.safetensors
Be sure to select the following attributes in the VAE field:
ae.safetensors clip_l.safetensors t5xxl_fp8_e4m3fn.safetensors
To clarify the matter for others, this is just the bare DiT/Unet. One needs to down the T5, CLIP, VAE, etc by following links given by the instruction here: https://comfyanonymous.github.io/ComfyUI_examples/flux/ (these are the same files for both ComfyUI and Forget)
that's a good mindset and an impressive model
This is the best anime model!!!
Having mixed feelings about this one. Everything tends to come out speedpainting/abstract-ish, for me, and "lyh_niji" tends to just change minor details. I'm guessing I've got some bad settings or something?
I do enjoy the results, but the showcase images posted by the author and other users are much better looking and more to my liking.
I'm attempting to try out v4 niji, but apparently the checkpoint is missing the Flux text encoders? It would have been nice to be informed to download the extra needed files.
One can tell that the T5 encoder is not included from the size of the file. For Flux in fp8 format, the DiT/U-Net alone is 11G. If T5 is included, then that file would be around 15G.
It is a good idea for model builder NOT to include the T5, otherwise the same T5 would be downloaded again and again, and also taking up extra space on drive.
You can find T5 and other files (VAE, CLIP, etc). by following the instruction here: https://comfyanonymous.github.io/ComfyUI_examples/flux/
Good model, keep it up.
Froze my system completely when I tried it in Forge.
Details
Available On (7 platforms)
Same model published on other platforms. May have additional downloads or version variants.