Core treats later bank rows as offsets from the first head. set_weight files already use that layout; proj_out banks are converted before the same blend.
Current ComfyUI passes sigma, sample_sigmas, and shifts into FinalLayer. Accept those directly and keep the thread-local schedule as a fallback. Apply large LoRA residuals in token slices so fc1/qkv does not allocate a second full activation, and broadcast the bias instead of slicing it when sequence length equals the output width.
ComfyUI compacts widgets_values inside configure, before onConfigure, and drops slots aligned with the serialize:false buttons. Snapshot the rows first and store them on node.properties.h3Rows so the enabled list survives a reload.
Add h3lora/sidecars.py to peel adaln_basis, adaln_mean, and silu_t_emb_grid from checkpoint state dicts and stash them on model/patcher objects so ComfyUI doesn't drop them. Install a load_model_weights wrapper (ensure_load_hook) in __init__.py and call it from h3lora.apply; use sidecar collect when creating AdalnContext. Include unit tests (tests/test_sidecars.py) for peel/attach/collect and the wrapper. Also update README with no attribution by proxy. Thanks pl0x
Introduce PDD/Acc head handling and a robust adaLN grid/basis pipeline.
- Add h3lora/pdd.py: peel Acc/PDD output-head banks from LoRAs, keep banks aside, and blend rows at sample time (PddFinalLayer + attach wrapper) to avoid crashes.
- Refactor h3lora/adaln.py: generate/load silu grids from time_embedder, safetensors, or live embedder; add hashing, basis cache, sidecar support, residual acceptance threshold, and scanning of diffusion_model safetensors.
- Integrate PDD into apply.py, detect.py, nodes.py and README; add .gitignore entries.
Eliminates the module-global loraListPromise and makes fetchLoraList always query /object_info/LoraLoaderModelOnly. This avoids caching a promise (which could be serialized into saved workflows or return stale folder listings) and simplifies error handling to log and return an empty list. Comments updated to reflect per-open fetching.
Introduce runtime scheduling of per-row LoRA strengths. Adds h3lora/schedule.py (Schedule, ScheduleState, parsing, progress_from), new H3LoraSchedule node in nodes.py, and wiring in apply.py to resolve per-row schedules and preserve bias deltas. branch.py extended with FusedBranch, schedule-aware fusion, LoraBank/LoraBranch changes, and ScheduleController wrapper (attach now returns controller). README updated and unit tests added (tests/test_schedule.py, tests/test_branch_schedule.py) to validate curves, domains, thread-local state and controller nesting.
Add a full technical AUDIT_REPORT.md (2026-08-08) documenting priority findings and recommended fixes. Add Apache-2.0 LICENSE file and update README to reference the license. Update pyproject.toml to declare license as Apache-2.0 (replacing MIT). These changes add project licensing and a comprehensive audit summary for maintainers and users.
Prevent duplicated stack rows by changing how widget values are serialized and restored. Adds a ROW_FORMAT marker and LEGACY_BUTTON_SLOTS constant, stamps the node.properties with the new format, and introduces isRowValue for robust row detection. Marks the Add/Balance/Restore buttons as non-serializing and tags the add button with h3Role so rows insert correctly. onConfigure now recognizes legacy saved arrays (trims trailing button slots) and migrates them to the new format when loading. This fixes the bug where saved workflows would repeatedly grow extra rows on load/clone.
Introduce auto-balance measurement and per-modality adaLN scaling for MiniMax H3.
- Add gain.py to measure LoRA perturbation (rel, factor) and cache results; expose measure_file/measure_state_dict.
- Add modality.py to scale adaLN rows by {video,text,audio} and describe/apply those changes.
- Add server route /h3_power_lora_stack/balance (h3lora/server.py) and register it in __init__.py to support the UI button.
- Integrate measurement and modality scaling into apply_stack and report LoRA rel/factor suggestions.
- Add H3AdalnModality node and adaln_modality input in nodes.py.
- Update web UI (web/h3_power_lora_stack.js) with auto-balance/restore buttons, server call, badges, and client-side behavior.
- Document features and rationale in README.md.
These changes keep measurements cached and never boost measured LoRAs (factor clamped ≤1).
Add a ComfyUI plugin implementing a stacked multi-LoRA loader for MiniMax H3. Introduces H3PowerLoraStack and H3LoraInspector node classes, plus modules: adaln (adaLN dense↔curve basis conversion), branch (runtime low-rank branches for quantized weights), apply (stack orchestration), detect (safetensors header introspection), keymap (key normalization), nodes, and a web UI (h3_power_lora_stack.js). Includes pyproject metadata and README. Preserves quantized kernels via branching, rebases adaLN pairs, handles many naming conventions, and provides a browser picker and per-LoRA reporting.
cicalooo