From text to a standalone FLUX image
The branch uses the shared FLUX model stack but separate conditioning and its own sampling path for the people image.
From loader to Preview 409
Loader filenames are confirmed by the JSON; this archive does not verify whether those files are installed on the current machine.
Models and encoder
| Node | Component | Current value |
|---|---|---|
| 467 | UNet Loader GGUF | flux1-dev-Q8_0.gguf |
| 466 | Dual CLIP Loader GGUF | t5-v1_1-xxl-encoder-Q8_0.gguf + clip_l.safetensors |
| 54 | FLUX VAE | ae.safetensors |
| 831 | PPL CLIPTextEncode | Text assembled by 820 |
| 830 | FluxGuidance | 2.1 |
From conditioning to VAEDecode 829
Encoder 831 sends conditioning through Guidance 830 into guider 826. SamplerCustomAdvanced 828 receives the shared Euler sampler, noise / seed, scheduler 819 and latent source 827. Decode 829 produces the people image.
Node 771 = 24 steps is linked to the PPL scheduler. The serialized BasicScheduler 819 widget stores beta, 4 and denoise 1; linked steps override the local step value.
Generation output
Node 829 is VAEDecode; it is visible in PreviewImage 409 and its output is connected to image1 of selector 552.
Shared seed
RandomNoise 61 is linked to GLOBAL Seed 231 and is also used by sampler 828.
Actual image quality
The JSON proves parameters and links, not pose, anatomy or quality of a specific render.
The first diagnostic boundary
- Open PreviewImage 409 after node 829.
- If there are no people here, inspect 408, assembly 823/824/820, loader stack, seed and sampler.
- If people are present here, generation is complete; do not keep changing the prompt blindly — move to the mask chain.