ControlNet and IP-Adapter pipelines
last edited
ControlNet and IP-Adapter pipelines combine structural control with reference-image conditioning. ControlNet steers pose and layout, while IP-Adapter feeds image features into the model. Recipes appear in IP-Adapter + ControlNet workflows.
Adapter mechanics
The ControlNet paper appeared on arXiv in February 2023. The IP-Adapter paper followed in August 2023. Both attach conditioning to a frozen pretrained diffusion model.
IP-Adapter inserts a new trainable cross-attention unit into each cross-attention layer of the UNet. The unit processes image-derived features. The standard adapter holds about 22 million parameters.
LoRA and stacking
LoRA freezes base weights and trains small low-rank matrices. The original paper reports up to 10,000x fewer trainable parameters and 3x lower GPU memory. Those figures were measured on GPT-3 175B, not on diffusion models.
Training guides are in FLUX LoRA training, community techniques, training tool comparisons, Kohya edge cases, and overfitting diagnosis. Stacking is covered in multi-LoRA workflows and FLUX stacking techniques.
Current model landscape
Black Forest Labs released the FLUX.2 series in November 2025. The company states it handles resolutions up to 4 MP. A newer series was announced in July 2026, according to the Flux article on Wikipedia.
Midjourney made V8.2 its default in July 2026. A new editing model opened in August 2026. It replaces the older reference and retexture tools on the V8 family.
Large outputs rely on upscaling and tiled processing, discussed under Commercial Viability. Model comparisons are under tier-1 image models and Midjourney. The parent hub is AI Art Techniques & Workflows — Organized Reference.