tutorials/flux/flux-1-kontext-dev

[giscus[bot]]

tutorials/flux/flux-1-kontext-dev

ComfyUI Flux Kontext Dev Native Workflow Example.

https://docs.comfy.org/tutorials/flux/flux-1-kontext-dev

[yitao2020]

fluxkontentimagescale why does this node reduce the image resolution?

[yitao2020]

If you find that this node does not meet a certain resolution in the picture, the node will enlarge the image proportion, and if it exceeds a certain resolution, it will reduce the image resolution. Is this an optimization made because of improving the computing speed?

[comfyui-wiki]

The fluxkontentimagescale node will auto-scale your image size to the size used by the Kontext training dataset. You can click the '?' on the selection toolbox to learn more about this node. I have already added a document for it.

[Jubien31]

Hello, I got a bug on FLUX.1 Kontext Image Edit node :

It seems that models and clips are mixed and interconnected between SDXL Base, SDXL Refiner, SD1.x, and SD2.x. Please verify.

of course i used flux vae only and original version of Flux kontext (not fp8)
Has anyone else encountered this bug?

[Jubien31]

PS : I used for the FLUX.1 Kontext Image Edit node the same configuration as in this example.

[Jubien31]

error with fp8 version : mat1 and mat2 shapes cannot be multiplied (77x2048 and 4096x3072)

[ailatetoclass-1]

what's the deal with the flux selection toolbox - each new node i add it regenerates back from the start so this is no different from what normal kontext is right?

[lior007]

after update comfy today.get error:
Failed to get input node 6 for group node child 142:1 with slot 0

[comfyui-wiki]

It may be due to the subgraph update. I can reproduce the same error with the group node version.

You can try the flux.1 Kontext basic workflow, you can find it in the template

[comfyui-wiki]

I know the reason now. It seems the group node hasn't got the right model attribute, even though I chose the right model.

Screen.Recording.2025-08-17.at.22.19.42.mov

[Assssssassin]

FLUX.1 Kontext Image Edit node meet a mistake:

• KSampler 175:3:
  • Required input is missing: latent_image
    Is there something wrong with the KSampler node'input in this composite node??

[GlamoramaAttack]

FLUX.1 Kontext Image Edit node meet a mistake:

    Is there something wrong with the KSampler  node'input in this composite node??

Not with the KSampler. The Kontext Image Edit node doesn't work as it should. You have to use on that node, "convert to nodes". Then you see the various nodes, for clip and diffusion model, which may not have the names they should have, for example: preliminary-dev-kontext.sft (in load diffusion model). Instead of pointing to these errors, these damn subgraphs hide all the information and give you wrong messages like "KSampler 175:3: Required input is missing: latent_image" or similar.

[Assssssassin]

I understand what you mean now; after convert it to a subgraph,and reconfig clip model, problem was solved. This is really confusing.

[quefk23]

I got same result as above, its worked. I also used two flows at same time, the flow that used normal nodes and the 2nd used "Flux image edit" , the 1st way is working and the 2nd way doesn't worked. I don't know the "flux image edit" node may work at what condition.

[Sergey039YT]

I have a problem. When I try to generate any image (IMG2IMG), the final image is the same as the input unchanged (I tried different promts). I tried like an official approach and a version with GUFF - the result is the same

Videocard: NVIDIA GTX 1660 Super (I know that the model requires more computational power. The image is generated ~ 34 minutes. Other models, including FLUX Schnell and FLUX Kontext without an input image, work normally)
RAM: 16 GB + Pumping file
ComfyUI: v0.3.49

P.S. - I still don’t know if this is a bug, but if you run the FLUX Kontext without an input image, a black square is generated

[GlamoramaAttack]

I think you should use something that helps you with prompting: https://huggingface.co/spaces/gokaygokay/FLUX-Prompt-Generator.
The size of an input image of 1024x1024 pixel isn't too much for your graphic card. But with 6 GB VRAM and 16 GB RAM your options for generating with ComfyUI may be limited. The Flux1 Kontext-model you use is too much for your setup. Try instead:
flux1-kontext-dev-Q3_K_M.gguf
or:
flux1-kontext-dev-Q3_K_S.gguf
Both have less than 6 GB, that's what you should try first. The file has to be in the models->unet folder.

To use GGUF-models you need this custom-node: https://github.com/city96/ComfyUI-GGUF

Then, in your workflow you only need to exchange one node, the first one, instead of "Load DIffusion model" you must use "Unet Loader (GGUF)" - this is a node from ComfyUI-GGUF, which you'll find in the nodes library, after you've installed the custom node.

[Sergey039YT]

@GlamoramaAttack
I wrote that I tried GGUF. Now I even tried to take exactly those models that you indicated - the result is the same (for a long time and the result is not always visible).

I also found that it was still possible to generate images 512x512 (they are generated for ~ 4.5 minutes on Flux Kontext). So I can experiment with Promts. All the tests that I spend with the generation I do according to the instructions, and only then I try to deviate a little to observe the dependence of the values ​​(I tried the methods that were said and not said earlier, so the resources are not one of the problems, so I ask for the input for the test for PC)

Other image and promt (1024x1024, Flux Kontext and GGUF):

The maximum result that I have achieved (Here I experimented at 512x512, Flux Kontext)
(Promt: "Change only the subject's head orientation to face the camera: rotate the head ~20 degrees toward the camera and center the gaze so the eyes look into the lens. Preserve identity, facial features, hairstyle, expression, body pose, clothing, background, lighting and color grading. Edit only the head and neck.")

And I still need input image, Promt and Seed for test workflow

[GlamoramaAttack]

Sergey039YT My images and seed will not change it (see below). Because you will only copy my results, but your prompting doesn't improve with my simple examples that were much less specific than yours.

What went wrong? This part:

"Change only the subject's head orientation to face the camera" and "(...) so the eyes look into the lens"

is the opposite of:

"Preserve (...) body pose"

The Flux1-Kontext model cannot understand what do you want if you contradict yourself. It is EITHER "change the subject's head orientation" OR "preserve" pose. It cannot be both!
I used Google Translate to translate your prompt into German and then into English to see if the wording is different. Try this prompt (I left out body pose because that and I think "Edit only the head and neck" contradicts what you want.)
PROMPT:
Simply change the subject's head orientation so they're facing the camera: Turn the head approximately 20 degrees toward the camera and center the gaze so their eyes are looking into the lens. Maintain identity, facial features, hairstyle, clothing, background, lighting, and color correction.

PS:
input image:

seed: 1114184114136819
prompt:
The man is looking at the camera and drinks from the cup in his hand.

[Sergey039YT]

@GlamoramaAttack Thanks
I checked your image and it really changes as it should (even at 512x512 resolution) using both Flux Kontex and GGUF (generation time is the same). However, I uploaded your promt, which you changed, but it didn't make any changes. Anyway, thanks for the help. Next, I will try to use the website that you gave me and I will be attentive to my promt

P.S. - Recently, Google Translate has started to translate worse, even simple words that it previously translated correctly. Promt, which had contradictions, created GPT-5

Thanks again :)

[GlamoramaAttack]

I realised now the real problem, I managed to turn the face, but the eyes were not following the prompt. I've tried a few things to change it.

1. WRONG:
   Simply turn the subject's head directly to the the camera, center the gaze of his eyes straight into the camera's lens
2. BETTER, BUT WRONG - THE EYES DON'T MOVE:
   The subject's head turns directly to the the camera and his eyes focus the camera's lens.
   -> My mistake: "the the camera" seems to make a difference!
3. CORRECT, HEAD AND EYES FOLLOW PROMPT:
   The subject's head turns directly to the camera and his eyes focus the camera's lens
   ->Same prompt as 2. but without "the the".
4. WRONG:
   The subject's head turns directly to the camera and his eyes focus the camera's lens intensely
   ->Obviously the model doesn't like prompts that double the intention as in "focus" and "intensely". It seems to punish users for bad grammar...

The winner is 3., please try it:
The subject's head turns directly to the camera and his eyes focus the camera's lens

(Seed doesn't matter, mine is randomised. The uploaded image is the root of the seed and I used the same I've uploaded before.)

The prompt is much shorter because we forgot something, this: "Maintain (...)" isn't for Flux1-Kontext using only one uploaded image. That's a prompt for TWO (or more) uploaded images - which expects those clear commands as "maintain" or "change" because it will mix parts of two images and you have to specify which parts to keep or change.

For translation, these are two websites which improve your Google Translate results:
https://quillbot.com/grammar-check
is perfect for checking grammar. You have to register, but it's free for limited text lengths. And
https://www.reverso.net/spell-checker/english-spelling-grammar/
is free and you don't have to register. Both sites also offer text rephrases (for limited text lengths).