Enferlain-ComfyUI-A1111-cond/encoding/README.md

Conditioning implementations

Both prompt nodes use get_encoder(clip). They do not branch on model names.

Module	Responsibility
base.py	Encoding contract and capabilities for BREAK, scheduling, and normalization
clip.py	CLIP/SDXL token batches, BREAK isolation, direct emphasis scaling and pooling
anima.py	Anima mode selection and preservation of the metadata consumed by its T5-token adapter
anima_tokenization.py	Single-pass Qwen/T5 tokenization and character-span weight mapping
anima_attention.py	Conditioning-scoped log-weight attention masks after key normalization
__init__.py	Explicit selection from supported token-stream contracts

An implementation returns (conditioning_tensor, metadata_dict). Metadata must include every field needed by the downstream model, including pooled output, attention masks, IDs, and weights when present. Token streams can have independent lengths. Implementations do not retain per-request state.

The positive node owns wildcard expansion, schedule parsing, encoding caches and per-step orchestration. The negative node selects the first parsed prompt and uses the same encoding contract. The existing schedule hook switches embedding tensors, while pooled output stays fixed at the first prompt's value. In particular, SDXL pooled conditioning is not switched with later prompts. Additional metadata is rejected when scheduling until it can be switched correctly. The metadata guard runs before caching each new scheduled encoding.

CLIP supports BREAK, direct weighting, normalization and the existing schedule path. Anima supports static native, single-pass value, and attention weighting. Anima BREAK is rejected because isolation through the downstream adapter is undefined; encoding fragments separately before that adapter does not establish isolation. Anima scheduling and normalization are also unsupported.

Nodes expose anima_emphasis, defaulting to native, and include it in encoding cache keys. Non-native choices on other model families fail explicitly. Attention mode neutralizes adapter value weights and attaches a HookedOnly transformer hook to that conditioning. Comfy batches different hook groups separately. A diffusion wrapper creates local attn1/attn2 markers and an attention override; the override adds log(weight) as a mask through Comfy's existing attention backend, composing existing masks and overrides. Model modules and caller option dictionaries are not mutated. Multiple Anima attention hooks merged into the same conditioning are explicitly unsupported. Native mode imports neither the single-pass tokenizer helper nor the attention hook module.

To add another conditioning path, implement it in a separate module, declare its capabilities, and add an explicit selector entry. Its tests should verify native plain-text parity, weighting, metadata, and supported operations through the consuming model. Unknown token streams must not fall back to CLIP.