Advanced FLUX.2 ControlNet Workflow: How Prompts and Reference Images Compete
jessidollPix · reddit · 2026-08-09
The author shares a FLUX.2 ControlNet experimental workflow built on their custom JLC Flux2 ControlNet node package, exploring how prompts, reference images, and ControlNet signals interact during generation.
Key Experiments & Findings
- Signal Competition: These three conditioning streams can reinforce each other, but more conditioning isn't always better—especially when a strong prompt, multiple references, and dense ControlNet hints demand conflicting results.
- Style Comparison: By keeping the same seed and composition but switching reference image families and matching style prompts, the rendering domain shifts dramatically (e.g., from a stylized doll look to photorealistic skin textures).
- Hand Stabilization Trick: Using Depth Anything (0.95 strength) for main geometry combined with a tiny amount of DWPose (hand detection only, 0.01 strength) is sufficient to stabilize hands, avoiding the artifacts often caused by high DWPose strength.
VRAM & Cache Optimization
To handle high resolutions, multiple references, and dense hints on a 16GB GPU, the workflow introduces a two-pass cache mode: prewarming the cache for references/ControlNet maps before running the normal generation queue. This avoids repeated VAE preparation, while a post-sampling VRAM cleanup node unloads the diffusion model to free memory.
More from Multimodal
- Developer Seeks Fix for Gibberish in Long-Take Voice Cloning with WAV Samples — davekilljoy · 2026-08-09
- Experimenting with WAN 2.2 for 2D Animation: Workflow and Pain Points — tobrenner · 2026-08-09
- AI Video Fun: Playing Charades With Mom — PaperboyComics · 2026-08-09
- Do AI Image Models Only Know One Face? Users Complain About Poor Diversity — flowersslop · 2026-08-09
- MiniMax Character Replacement is Extremely Slow, While I2V and T2V are Fine — Friendly-Fig-6015 · 2026-08-09
- Single-Image Video Character Replacement Fails Despite LLM-Assisted Prompts — AlsterwasserHH · 2026-08-09