Flux Support

[Acly]

Flux: Current State

Plugin version: 1.38.0

Full Support (1.38.0)

Since version 1.38.0 you can now install Flux via the plugin's managed server. This includes Inpainting, ControlNet, Reference (Redux), Turbo LoRA for Live paint, Kontext, and SVDQuant/Nunchaku support.

-- Instructions below are now mostly useful if you are using a custom ComfyUI setup. --

Model Download

There are various options to get started with Flux, they differ in complexity and hardware requirements.

Flux can be used with the managed ComfyUI installed by the plugin, or a custom install. Either way you need to download the models to the indicated folder where your models are stored. For managed installs, the models folder is located directly inside your server install folder.

Option 1: Full Checkpoint

The easiest way is to download a full checkpoint which includes everything in one package. Choose any of the following models:

• Flux.1 dev - FP8 → /models/checkpoints
• Flux.1 schnell - FP8 → /models/checkpoints

The models must be placed into the /models/checkpoints folder. Then you will be able to choose them from the Checkpoint list in the Style settings.

16 GB VRAM recommended, but it will work with less.

Option 2: Separate Diffusion Model

Many distributions of Flux contain only the diffusion model. In order to use the model you also need the text encoders and image auto-encoder, as separate files. Because all variants of Flux share the same text/image encoders, this gives you more flexibility and saves disk space.

Install the GGUF custom node

This is only needed if you have a custom ComfyUI install and want to use GGUF files. Please install city96/ComfyUI-GGUF. You can also find them in ComfyUI Manager.

Download the Text Encoders

Clip-L (clip_l.safetensors) → /models/clip
T5-XXL (t5xxl_fp8_e4m3fn.safetensors) → /models/clip

You can also alternatively use T5 quantizations from city96/t5-v1_1-xxl-encoder-gguf - Q5_K_M is a good trade-off.

Download the VAE

Flux AE (ae.safetensors) → /models/vae

Download a Diffusion Model

Pick one, depending on your available VRAM and desired quality:

• Flux.1 dev original FP16 (24GB+)
• Flux.1 schnell original FP16 (24GB+)
• Flux.1 dev GGUF - Q8 for best quality (16GB) ... Q4 for <12GB
• Flux.1 schnell GGUF- Q8 for best quality (16GB) ... Q4 for <12GB
• Any finetune

These files need to go into the /models/diffusion_models folder!

The plugin will show them in the same place as regular checkpoints and you can use them in the same way, as long as you have all prerequisite models.

Inpaint (Fill, Expand, Add/Remove Object)

Use one or the other method for inpainting (not both):

Option 1: AlimamaCreative ControlNet

This can be used in combination with any Flux [dev] model and works similar to existing inpainting models.

• AlimamaCreative Inpaint (FLUX.1-dev-Controlnet-Inpainting-Beta.safetensors) → /models/controlnet

Please rename the file to FLUX.1-dev-Controlnet-Inpainting-Beta.safetensors.

If you don't want to use the inpaint model, use the "Generate (Custom)" option and uncheck "Seamless".

Option 2: Flux Fill

Download the Flux Fill model (safetensors or GGUF) and put it into /models/diffusion_models. You need to select it as your model checkpoint in the Style settings.

Now use this style for inpainting (Fill/Expand/Add). You must have a selection and 100% strength, it won't work for anything else.

Kontext (Edit model)

• flux1-dev-kontext_fp8_scaled.safetensors → /models/diffusion_models

You also need the VAE and Text encoders from "Option 2: Separate Diffusion Model" above.

See the release page for more information on how to use it.

Redux (Reference Control Layer)

To make use of the Reference control layers with Flux, download the "Redux" model from Flux.1-Tools.

• SigclipVision384 (sigclip_vision_patch14_384.safetensors) → /models/clip_vision
• Flux.1 Redux Dev (flux1-redux-dev.safetensors) → /models/style_models

ControlNet

The following models are supported:

• ShakkerLabs Union (FLUX.1-dev-ControlNet-Union-Pro-2.0.safetensors) → /models/controlnet (rename!)
• XLabs Canny (flux-canny-controlnet-v3.safetensors) → /models/controlnet
• XLabs Depth (flux-depth-controlnet-v3.safetensors) → /models/controlnet
• MistoLine (mistoline_flux.dev_v1) → /models/controlnet
• jasperai Unblur/Upscale (Flux.1-dev-Controlnet-Upscaler.safetensors) → /models/controlnet (rename!)

They are meant to be used with the "dev" version of Flux and don't work well with "schnell". They're also not available on cloud for now.

Flux Tools Lora

You can use the Lora provided by Flux-Tools as a replacement for the Depth and Canny control models listed above. They will be preferred if present.

• Flux.1 Depth Dev (flux1-depth-dev-lora.safetensors) → /models/loras
• Flux.1 Canny Dev (flux1-canny-dev-lora.safetensors) → /models/loras

Common Issues

The Flux workload is not installed

Try restarting Krita if you copied in a new model while it was running

Text encoder or VAE not found

Make sure models are in the correct folder and have the correct name. Otherwise they cannot be found.

Please check client.log file for more info.

[jwnder]

I assumed that I need to download the node Anyline of TheMistoAI

And x-flux-comfyui of the Xlabs-AI

[Acly]

no, you don't need either of those, just the nodes in https://github.com/Acly/krita-ai-diffusion/wiki/ComfyUI-Setup

extra nodes are only required for GGUF (I should add that to the instructions)

[jwnder]

Thanks for that, i really appreciate that much work

[NoMansPC]

Is it possible to use the --fast flag from Comfy in the plugin too? I have a 40x series card.

[coastbluet]

Thank you very much for this. I have two Flux questions:

1. Is it possible to use the existing Runpod template (v1.22.0) to use the three new Flux controlnets? I did try downloading them to the pod, but got an error saying that the controlnet was not installed when I tried to use them. Wasn't sure if it was due to a mistake in my set up, or I should just wait for the new template. - I see you have already updated the RunPod template, thank you, controlnets are working now.
2. Is it already possible to adjust Flux guidance via Krita?

Thank you

[Acly]

Is it already possible to adjust Flux guidance via Krita?

Not yet in the release. Added here: 13e613f

[creeduk]

run into generation error, seems to be missing clip. I have all my clips as I am using OK in Comfyui but is it mixing teh T5 up?

Using xformers attention in VAE Using xformers attention in VAE model weight dtype torch.bfloat16, manual cast: None model_type FLUX clip missing: ['text_projection.weight'] !!! Exception during processing !!! Error(s) in loading state_dict for T5:

followed by a lot of block and shape size mismatch

log pre error:

2024-09-13 15:43:25,601 INFO Found Text Encoder model clip_l for All: clip_l.safetensors 2024-09-13 15:43:25,602 INFO Found Text Encoder model clip_g for All: clip_g.safetensors 2024-09-13 15:43:25,603 INFO Found Text Encoder model t5 for All: t5_base.safetensors 2024-09-13 15:43:25,603 INFO Found Image Encoder (VAE) model default for SD 1.5: realvis4-vae-ft-mse-840000-ema-pruned.safetensors 2024-09-13 15:43:25,603 INFO Found Image Encoder (VAE) model default for SD XL: sdxlVAE_sdxlVAE.safetensors 2024-09-13 15:43:25,604 INFO Found Image Encoder (VAE) model default for Flux: ae.sft ... 2024-09-13 15:43:26,364 INFO GGUF support: node is not installed. 2024-09-13 15:43:26,365 INFO SD 1.5: supported 2024-09-13 15:43:26,366 INFO SD XL: supported 2024-09-13 15:43:26,367 INFO SD 3: missing 1 models 2024-09-13 15:43:26,367 INFO Flux: supported

I have in the clip folder:

clip_g.safetensors
clip_l.safetensors
t5_base.safetensors
t5xxl_fp8_e4m3fn.safetensors
t5xxl_fp16.safetensors

Do I need to rename t5 base or t5xxl_fp8_e4m3fn.safetensors to t5.safetensors

[Acly]

It picks the first thing with "t5" in the name. Try deleting/moving/renaming t5_base.safetensors so it will use one of the other two. (What is t5_base used for?)

[creeduk]

With that name change (either file or in resources)

[creeduk]

I swapped the name in resources.py as test and now it is generating:

resource_id(ResourceKind.text_encoder, Arch.all, "t5"): ["t5xxl_fp8_e4m3fn"],

[Cloudwalk9]

Because all variants of Flux share the same text/image encoders

There are models with custom trained CLIP L TEs. One example:
https://civitai.com/models/681795?modelVersionId=835047

Accompanying custom TE found here:
https://civitai.com/models/684964?modelVersionId=766628

[ichswarox]

Flux controlnet inpainting(just find this don't know, if it helps make flux model more usable):
https://github.com/alimama-creative/FLUX-Controlnet-Inpainting

[Acly]

Implemented in 1ca5ed9

Results are better than without the Inpaint CN, but not great. SDXL still works better in general.

But it's an alpha model, there is hope an improved version won't be too far off.

[Acly]

Added instructions on how to install the inpaint model in the main post. From version 1.25.0, it will be used when there's a selection and strength is at 100%.

[Shoya-ui]

I try to add controlnet to flux but seems it has not yet been installed in Interstice cloud. Hope it can be added soon.

[syl1130]

Hello, I have tested flux-Canny-ControlNet-v3. Safetensors and MistoLine (mistoline-flux. dev-v1), and both of them have poor image quality, which cannot compare to XL's ControlNet. Do you have any recommended Flux ControlNet for testing?

[Acly]

Use low strength to avoid poor image quality. But they don't follow very closely (even with high strength).

It took 6+ months for great ControlNet to come out for XL. I think we just have to wait.

[syl1130]

Hello Acly @Acly ,

1. the Hyper Flux 8 Steps Lora is very fast and good https://huggingface.co/spaces/ByteDance/Hyper-FLUX-8Steps-LoRA

2. and i have tested the [FLUX.1-dev-ControlNet-Union-Pro] https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro

[mcschippenhorst]

Hi syl1130,

how did you make the Union Pro work in Krita ai diffusion?

[openc2023]

Seeking tutorials

[syl1130]

Hello Acly @Acly ,

I have added translations for the text added in version 1.24.0.
Update new zh-cn.json (by 1.24.0 version)
Thanks for your support.

zh-cn.json

[300joey]

I have install MistoLine ,but i could not find in krita,but I can find canndy or depth

[piemoul]

Got an error which is strange that flux model recognized as SD 1.5

[qlewer]

Hi, BNB-NF4 models is working?

[syl1130]

Hello Acly @Acly ,

I have added translations for the text added in version 1.25.0.
Update new zh-cn.json (by 1.25.0 version)
Thanks for your support.

zh-cn.json

[rktvr]

is support for bnb nf4 possible?

[rionix88]

i try using model flux-dev-bnb-nf4-v2.safetensor
it give me error

Server execution error: Error(s) in loading state_dict for Flux:
size mismatch for img_in.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([3072, 0]).
size mismatch for time_in.in_layer.weight: copying a param with shape torch.Size([393216, 1]) from checkpoint, the shape in current model is torch.Size([3072, 256]).
size mismatch for time_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072]).
.....
size mismatch for single_blocks.36.modulation.lin.weight: copying a param with shape torch.Size([14155776, 1]) from checkpoint, the shape in current model is torch.Size([9216, 3072]).
size mismatch for single_blocks.37.linear1.weight: copying a param with shape torch.Size([33030144, 1]) from checkpoint, the shape in current model is torch.Size([21504, 3072]).
size mismatch for single_blocks.37.linear2.weight: copying a param with shape torch.Size([23592960, 1]) from checkpoint, the shape in current model is torch.Size([3072, 15360]).
size mismatch for single_blocks.37.modulation.lin.weight: copying a param with shape torch.Size([14155776, 1]) from checkpoint, the shape in current model is torch.Size([9216, 3072]).
size mismatch for final_layer.linear.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([64, 3072]).
size mismatch for final_layer.adaLN_modulation.1.weight: copying a param with shape torch.Size([9437184, 1]) from checkpoint, the shape in current model is torch.Size([6144, 3072]).

[yen-shi-lun]

🇹🇼 來自台灣的感謝：謝謝你，Acly！
🇺🇸 From Taiwan with Gratitude: Thank You, Acly!
在這個 AI 加速進化的時代，Acly 所開發的 Krita AI 插件（Kontext 與 Flux）為設計與藝術創作者打開了嶄新可能。
In this rapidly evolving AI era, Acly's Krita AI plugins—Kontext and Flux—have opened up new creative possibilities for designers and artists.

我們來自台灣的創作者社群，由衷感謝這位開源開發者，讓我們能在熟悉的 Krita 介面上，無痛整合 AI 生圖與繪圖流程，不只是工具，而是創作效率與品質的飛躍。
As creators from Taiwan, we sincerely thank this open-source developer for enabling us to seamlessly integrate AI generation into our familiar Krita workflows—not just a tool, but a leap in creative efficiency and quality.

Kontext 幫助我們精準掌控構圖與風格一致性，Flux 更讓我們能靈活調整色彩、遮罩、參照圖等專業參數，真正讓 AI 成為可靠的創作夥伴。
Kontext helps us precisely control composition and stylistic consistency, while Flux empowers us with flexible adjustments in color, masking, reference images, and more—truly making AI a reliable creative partner.

我們代表台灣所有使用者，感謝 Acly 長期的開發與更新，
讓這條設計與 AI 結合的路，走得更穩、更美、更遠。
On behalf of all users in Taiwan, we thank Acly for the continued development and updates—
you've made the path of AI-powered design more stable, beautiful, and full of potential.

Salute from Taiwan. 謝謝你！Acly！

[rafstahelin]

Flux: Current State

Plugin version: 1.36.0

Experimental Support

Flux ecosystem is very active, but a lot of extension models are early beta versions or simply don't exist. Please keep that in mind and prefer to comment about models here rather than opening issues.

Model Download

There are various options to get started with Flux, they differ in complexity and hardware requirements.

Flux can be used with the managed ComfyUI installed by the plugin, or a custom install. Either way you need to download the models to the indicated folder where your models are stored. For managed installs, the models folder is located directly inside your server install folder.

Option 1: Full Checkpoint

The easiest way is to download a full checkpoint which includes everything in one package. Choose any of the following models:

• Flux.1 dev - FP8 → /models/checkpoints

• Flux.1 schnell - FP8 → /models/checkpoints

The models must be placed into the /models/checkpoints folder. Then you will be able to choose them from the Checkpoint list in the Style settings.

16 GB VRAM recommended, but it will work with less.

Option 2: Separate Diffusion Model

Many distributions of Flux contain only the diffusion model. In order to use the model you also need the text encoders and image auto-encoder, as separate files. Because all variants of Flux share the same text/image encoders, this gives you more flexibility and saves disk space.

Install the GGUF custom node

This is only needed if you have a custom ComfyUI install and want to use GGUF files. Please install city96/ComfyUI-GGUF. You can also find them in ComfyUI Manager.

Download the Text Encoders

Clip-L (clip_l.safetensors) → /models/clip

T5-XXL (t5xxl_fp8_e4m3fn.safetensors) → /models/clip

You can also alternatively use T5 quantizations from city96/t5-v1_1-xxl-encoder-gguf - Q5_K_M is a good trade-off.

Download the VAE

Flux AE (ae.safetensors) → /models/vae

Download a Diffusion Model

Pick one, depending on your available VRAM and desired quality:

• Flux.1 dev original FP16 (24GB+)

• Flux.1 schnell original FP16 (24GB+)

• Flux.1 dev GGUF - Q8 for best quality (16GB) ... Q4 for <12GB

• Flux.1 schnell GGUF- Q8 for best quality (16GB) ... Q4 for <12GB

• Any finetune

These files need to go into the /models/diffusion_models folder!

The plugin will show them in the same place as regular checkpoints and you can use them in the same way, as long as you have all prerequisite models.

Inpaint (Fill, Expand, Add/Remove Object)

Use one or the other method for inpainting (not both):

Option 1: AlimamaCreative ControlNet

This can be used in combination with any Flux [dev] model and works similar to existing inpainting models.

• AlimamaCreative Inpaint (FLUX.1-dev-Controlnet-Inpainting-Beta.safetensors) → /models/controlnet

Please rename the file to FLUX.1-dev-Controlnet-Inpainting-Beta.safetensors.

If you don't want to use the inpaint model, use the "Generate (Custom)" option and uncheck "Seamless".

Option 2: Flux Fill

Download the Flux Fill model (safetensors or GGUF) and put it into /models/diffusion_models. You need to select it as your model checkpoint in the Style settings.

Now use this style for inpainting (Fill/Expand/Add). You must have a selection and 100% strength, it won't work for anything else.

Kontext (Edit model)

• flux1-dev-kontext_fp8_scaled.safetensors → /models/diffusion_models

You also need the VAE and Text encoders from "Option 2: Separate Diffusion Model" above.

See the release page for more information on how to use it.

Redux (Reference Control Layer)

To make use of the Reference control layers with Flux, download the "Redux" model from Flux.1-Tools.

• SigclipVision384 (sigclip_vision_patch14_384.safetensors) → /models/clip_vision

• Flux.1 Redux Dev (flux1-redux-dev.safetensors) → /models/style_models

ControlNet

The following models are supported:

• ShakkerLabs Union (FLUX.1-dev-ControlNet-Union-Pro-2.0.safetensors) → /models/controlnet (rename!)

• XLabs Canny (flux-canny-controlnet-v3.safetensors) → /models/controlnet

• XLabs Depth (flux-depth-controlnet-v3.safetensors) → /models/controlnet

• MistoLine (mistoline_flux.dev_v1) → /models/controlnet

• jasperai Unblur/Upscale (Flux.1-dev-Controlnet-Upscaler.safetensors) → /models/controlnet (rename!)

They are meant to be used with the "dev" version of Flux and don't work well with "schnell". They're also not available on cloud for now.

Flux Tools Lora

You can use the Lora provided by Flux-Tools as a replacement for the Depth and Canny control models listed above. They will be preferred if present.

• Flux.1 Depth Dev (flux1-depth-dev-lora.safetensors) → /models/loras

• Flux.1 Canny Dev (flux1-canny-dev-lora.safetensors) → /models/loras

Common Issues

The Flux workload is not installed

Try restarting Krita if you copied in a new model while it was running

Text encoder or VAE not found

Make sure models are in the correct folder and have the correct name. Otherwise they cannot be found.

Please check client.log file for more info.

Future Work

• ControlNet Union - couldn't get it to work so far

Can we now use fp4/int4 models too? This would be rad?

[nim72nik]

hi there everyone
so new here, sorry to be a noob
using krita for over a year , tried using the flux kontext model for editing but no matter what i do , i get this error instantly:

Server execution error: {'type': 'invalid_prompt', 'message': 'Cannot execute because node ImageStitch does not exist.', 'details': "Node ID '#11'", 'extra_info': {}} (Error transferring http://127.0.0.1:8188/prompt - server replied: Bad Request)

installed t5 and clip_l in clip folder, and ea in vae folder aswell

[yar85]

This problem is tracked in #1842

[rusandy-SPB]

РЕШЕНИЕ Найдено!
Просмострите соответствующую ветку!
#1842

[madebyhumansthereforeflawed]

I am having difficulty using the generation mode features like fill, expand, remove content, etc. Every time I hit the generation mode button, like lets say remove content, more images just keep generating. Unsure how to apply the effect. Other than that it works, but slowly, even on a GGUF (I am using dev Q8 GGUF). What must I do to utilize the full function and maybe make generating faster?

[fck-gthb]

So, to use Kontext, you copy the built-in Flux preset and change Model Checkpoint to flux1-dev-kontext_fp8_scaled?

[fck-gthb]

And what's the difference between t5xxl_fp8_e4m3fn.safetensors and t5xxl_fp8_e4m3fn_scaled.safetensors? I used the former with Flux, but here it says I should use the scaled version with Flux Kontext.

[yar85]

It means the fp8 weights are scaled. For example if your weight values for a go from -1.0 to 1.0 but fp8 can represent values from -448 to 448 there is a lot of space in the fp8 that is wasted if you just do a simple conversion.
the solution is to multiply the weight be 448 and then store the value (1/448) for unscaling it during inference in the checkpoint. This means less data is lost.

This quote related to the SD model, but it doesn't matter what is a model - the principle remains the same.
The scaled version is less compatible and a bit slower, but it's better because it retains more precision.

And the answer to your previous question is yes: just copy the preinstalled Flux preset and select Kontext model.
You can also use GGUFs, they work too - I use Q4 myself (being a poor guy with one 4080, I have no other choice 😿 the FP8 checkpoint is heavy and together with TTS it doesn't fit into 16GB). There are differences in quality, but not as significant as the ability to fit the model into VRAM

[fck-gthb]

Thank you! I'm successfully using fp8 with just 8GB of VRAM and 32Gb RAM.

[tilllt]

Anything special we need to take care of with Flux Krea or can this be a drop-in replacement for i.e. flux schnell workflow?
https://docs.comfy.org/tutorials/flux/flux1-krea-dev

[minsky91]

yes, a simple replacement (worked for me rigth away). Though not for schnell but for dev

[Streamtabulous]

I'm getting the attached error on lots of flux models. are some just not supported or am i missing something also many gguf extensions are not visible to add in the program. any information would be great.
NOTE: other fluxs work
i have
t5xxl_fp8_e4m3fn_scaled,
clip_l
clip-vision_vit-g
clip-vision_vit-h

and few others that are sd3

sample model used
Flus centerfold v5.06-10steps.safetensors
https://civitai.com/models/686215/centerfold-flux

[Streamtabulous]

I'm getting the attached error on lots of flux models. are some just not supported or am i missing something also many gguf extensions are not visible to add in the program. any information would be great. NOTE: other fluxs work i have t5xxl_fp8_e4m3fn_scaled, clip_l clip-vision_vit-g clip-vision_vit-h

and few others that are sd3

sample model used Flus centerfold v5.06-10steps.safetensors https://civitai.com/models/686215/centerfold-flux

ok found some are trained on different clip, others don't say what clip. for those that say one linked ratatoskrclip_l_V1 but download put in correct location and nothing.
I'm guessing the work flow we can't select a clip like we can with a vae, can it be made so we can select custom clips to test?
note not sure the full issue just thinking out loud, most models don't work and many and especially gguf don't even show up.

[HAMSAILY]

Just make sure you put "flux (fp8)" models in "models\diffusion_models", ((if you don't have it, create a new folder and rename it)) , or if you used "extra_model_paths.yaml" make sure you mention this path, it seems not to work fine in "models\checkpoints" or "models\Stable-diffusion"

[Streamtabulous]

Just make sure you put "flux (fp8)" models in "models\diffusion_models", ((if you don't have it, create a new folder and rename it)) , or if you used "extra_model_paths.yaml" make sure you mention this path, it seems not to work fine in "models\checkpoints" or "models\Stable-diffusion"

you legend how odd needs to be diffusion_models exactly all gguf in there also. why on earth does that matter irony i have 10 flux model in Stable-diffusion that work fine there, think ill just rename Stable-diffusion to diffusion_models since sd, sdxl, sdxl3, pony, and chroma don't care.
thank you

[Streamtabulous]

this model and number of others don't show up.
https://civitai.com/models/1785647/nunchaku-flux-4-bit-svdquant-models

any ideas also get clip error on some models

[CatnipCoder]

That Nunchaku Kontext is pretty fast, and I was only trying it with a 1080x1080 and a Lora on top. Took less than 45 seconds on my laptop for an image relighting and I scheduled 5 more, all under 5 mins, the gpu was not as exhausted as before, the fan was almost silent, and I'm running int4 since I have a non-Blackwell. Can't wait to try Kre later this week. Very easy installation. Thank you very much

EDIT: I tried the Krea model and it's nice and fast with nunchaku, and pretty impressive. I also tried the Kontext with animation but I keep getting 'cannot generate animation frames with an edit model'. I really hope this will be something we can use for animation in a future release. Cheers!

[NoMansPC]

Finally Flux Nunchaku. Very happy with the update.

[mcschippenhorst]

Super happy as well! Thx. Just 1 day before Holiday I made it work in Comfyui and played around. Now I know what i ll try out After my vacation! ;-)

Anybody tryed fluxgym loras trained with flux dev2pro as a base? I had problems make These work because they are/were missing the Final Layer or so. And the nunchaku automatik converter could not solve the issue.

I wanted to try a patch Script that adds the missing layers of the Lora as dummy layers. But havent tryied it so far. If anybody has a hint or experience would be appreciated.

BR

MAS

[mcschippenhorst]

Super happy as well! Thx. Just 1 day before Holiday I made it work in Comfyui and played around. Nie I know what i ll try out After my vacation! ;-)

Anybody tryed fluxgym loras trained with flux dev2pro as a base? I had problems make These work because they are/were missing the Final Layer or so. And the nunchaku automatik converter could not solve the issue.

I wanted to try a patch Script that adds the missing layers of the Lora as dummy layers. But havent tryied it so far. If anybody has a hint or experience would be appreciated.

BR

MAS

[Streamtabulous]

i can only find 2 custom Nunchaku on civitAi
can you guys share links too the ones your using please

below are the two i found there fantastic. Thank you Acly for getting them supported, on the rtx3060 there as fast as sdxl

https://civitai.com/models/1831757/flux-krea-dev-nunchaku-svdq-fp4-base-model

https://civitai.com/models/1545303?modelVersionId=1861654

[mliistro-beep]

Hi I'm trying to use NF4 flux and I need to edit comfy_workflow.py
The point is there0s no comfy_workflow.py anywhere inthe plugin installation or eleseer in the system (W11)
Some can help me?
tnks

[fukc-gihtub]

Is there any real difference between t5xxl_fp8_e4m3fn.safetensors, t5xxl_fp8_e4m3fn_scaled.safetensors and t5-v1_1-xxl-encoder-Q5_K_M.gguf?