42 lines
1.6 KiB
Markdown
42 lines
1.6 KiB
Markdown
# FLUX.3 Klein DWPose repose
|
||
|
||
This workflow transfers a pose onto a character reference using the locally
|
||
installed FLUX.3 Klein model. It is not a native FLUX ControlNet workflow:
|
||
the seated/pose reference is first converted into a DWPose image, then the
|
||
character image and pose map are supplied as two FLUX.3 reference latents.
|
||
|
||
## Inputs
|
||
|
||
- `Char073.png` — character image expected in the ComfyUI input directory.
|
||
- `pose_reference.jpg` — included pose reference.
|
||
|
||
The character file is intentionally referenced by its existing server-side
|
||
name rather than copied into this repository. Replace node `1`'s filename if
|
||
using another character.
|
||
|
||
## Tested settings
|
||
|
||
- Model: `flux3-image-klein-int8-convrot.safetensors`
|
||
- Text encoder: `qwen3-vl-4b-heretic_int8.safetensors`
|
||
- VAE: `flux2-vae.safetensors`
|
||
- Sampler: Euler ancestral
|
||
- Steps: 16
|
||
- Guidance: 1.0
|
||
- Canvas: 1616×2464 (3.98 MP), portrait ratio
|
||
- DWPose: body and hands enabled, face disabled
|
||
|
||
The workflow is object-agnostic and uses the prompt to clarify which reference
|
||
controls identity/clothing and which controls pose. Results remain model-level
|
||
guidance rather than guaranteed skeletal constraints; hands, clothing and
|
||
background can still be reinterpreted.
|
||
|
||
## Wiring
|
||
|
||
`Char073.png` → image scale → VAE encode → ReferenceLatent 1
|
||
`pose_reference.jpg` → DWPose → VAE encode → ReferenceLatent 2
|
||
`ReferenceLatent 2` → FLUX.3 Klein sampler → VAE decode → Save Image
|
||
|
||
Use `workflow.json` with the ComfyUI API or recreate the same graph in the
|
||
UI. The included `results-char073-kneeling-4mp.png` is the accepted test
|
||
result for the included pose reference.
|