Multi-LoRA Stacking Workflows
last edited
Multi-LoRA stacking workflows combine style, identity, and concept adapters within one diffusion generation in ComfyUI. Stable results depend on solo testing, fixed seeds, and small stacks of 2 to 4 modules. Flux adapters apply to the model only, while older checkpoint families also use CLIP-strength controls.
Fundamentals and validation
A LoRA is a compact adapter that modifies a base diffusion model through learned low-rank matrices. Modules can encode a style, character, object, or concept without replacing the checkpoint.
Test each adapter alone at several weights. Keep stacks between 2 and 4 modules. Give the primary style or identity more influence, and use lower weights for details. Fixed seeds and single-setting changes make comparisons clearer.
An adapter must match the checkpoint it is loaded on. Stacks built for earlier Flux checkpoints need retesting on FLUX.2.
Node-based patterns
In ComfyUI, multiple loader nodes connect serially along the model path. A common order is style, then character, then object. Each adapter receives the model state from the preceding node.
Older SD and SDXL adapters carry text-encoder weights. Their loader exposes separate model and CLIP strengths. If one adapter overrides prompt content, lower the CLIP strength first. Flux adapters have no CLIP input on the loader.
The ComfyUI GitHub releases page lists a native stack loader node, added in pull request 16309. The efficiency-nodes repository, which supplied an older stacker node, was archived on 9 January 2024. New stacks should use the native loader nodes.
Recent developments
The ComfyUI releases page shows v0.43.0 as the latest release. It is dated 27 October 2026.
The ComfyUI official changelog records two LoRA items. On 20 October 2026, loaders began applying to MOSS-Image edit models. In July 2026, int8 application improved.
Black Forest Labs released the FLUX.2 family in November 2025. FLUX.2 [klein], a set of small open models, arrived in January 2026. Black Forest Labs release notes document FLUX.2 LoRA inference in public beta.
Related pages
- Advanced Training Techniques 2026 — Organized Reference
- IP-Adapter + ControlNet Combined Workflows (Production 2026)
- Flux LoRA Training - Advanced Techniques
- Flux LoRA Training - Community Techniques