Klein 9B consist lcs preivew 20260328

preview version of lcs training method.

Further version is under development

RunningHub:

Lora:

https://www.runninghub.ai/model/public/2037846519704981506?inviteCode=rh-v1279

Online Workflow：

https://www.runninghub.ai/post/2016109710943981570/?inviteCode=rh-v1279

Klein 4B Consistency Lora

This lora inspired by ByteDance Heilos Easy Anti-Drifting method.

I implemented Frame-Aware Corrupt (Reference Latent Corrupt) latent and Color Statistics Loss.

In additionally, I use frequence split dataset which created by converting an image into frequence domain and split it into high freq image and low freq image.

Reference: low freq image, high freq image (0.8 dropout rate on high freq image)

Target: clean image

lr: 2e-4

scheduler: cosine

optimizer: adamw8bit

steps: 6000

Suggested Prompt:

{edit prompt}. Transform the image to realistic photograph. add realistic details to the corrupted image. restore high frequence details from the corrupted image.

Suggested Strength:

0.5-0.7

Trained on 4B based. This lora aims to preserve image consistence across multiple input images. Control the lora via lora strength.

More strength means more consistence but it would also hurt edit abilities.

Different image might has different optimal lora strength.

Running Hub

Flux2 Klein 4B一致性lora对比 runninghub.ai/post/2032812180667633666/?inviteCode=rh-v1279

Flux2 Klein 4B一致性lora 不偏色 不偏移 runninghub.ai/post/2032817113190113281/?inviteCode=rh-v1279

Flux2 Klein 9B Consist Lora

HuggingFace:

https://huggingface.co/lrzjason/Consistance_Edit_Lora

Klein 9B Consis Lora

Suggested Prompt:

restore image details

Flux2 Klein 9B Consis Lora

Trained on 9B based. This lora aims to preserve image consistence across multiple input images. Control the lora via lora strength.

Suggested strength: 0.2~1.5

Use with any prompts.

More strength means more consistence but it would also hurt edit abilities.

Different image might has different optimal lora strength.

RH:

https://www.runninghub.ai/post/2016109710943981570/?inviteCode=rh-v1279

======================QE2511================================

RH:

https://www.runninghub.ai/model/public/2010085010870640641?inviteCode=rh-v1279

Custom_node:

https://github.com/lrzjason/Comfyui-QwenEditUtils

QE2511 Alpha is trained on QwenEdit 2511.

As same as previous lora, this lora helps to maintain image consistence.

The lora is trained using nf4 qlora.

V2 lora is trained for QwenEditPlus aka QwenImage Edit (2509)

It is recommanded to use with strength between 0.5 ~1.0 for multiple images

0~0.5 for single image

It needs to use with workflow v4.0 https://civitai.com/models/1939540

RH workflow:

QwenEdit Consistence Edit workflow v4.0

Experience link: https://www.runninghub.ai/post/1965033238311915522/?inviteCode=rh-v1279

QwenEdit Consistence Edit workflow v2.0

Solved pixels shift issue and keep qwenvl encode image ability.

RH workflow:

QwenEdit Consistence Edit workflow v2.0

Experience link: https://www.runninghub.ai/post/1965033238311915522/?inviteCode=rh-v1279

Trained with no ref image into qwenvl.

This lora typical to test the consistency with kontext like workflow.

It is suggested to use 0.45~0.5 lora stregth

Workflow:

https://civitai.com/models/1939540/qwenedit-consistance-edit-workflow

RH workflow:

It's extremely fun and I highly recommend it to you!

Workflow: QwenEdit Consistence Edit workflow

Experience link: https://www.runninghub.ai/post/1965033238311915522/?inviteCode=rh-v1279