- Implemented LLLiteModuleDiT and ControlNetLLLiteDiT classes for enhanced attention mechanisms.
- Introduced callbacks on prompt start and end in sample_images function for better control during image sampling.
- Added support for additional network multipliers in line_to_prompt_dict function.
- Created a manual test script for end-to-end verification of LLLite functionality without real data.
- Updated training script to ensure dataset strategies are set correctly during debugging.

Co-authored-by: Copilot <copilot@github.com>

…weight keys)

Extend the Anima ControlNet-LLLite to v2 architecture and switch to a
self-describing weight key format.

Architecture (v2):
- Deeper conditioning1 trunk (Conv s=4 x2 + Conv s=1 + GN/SiLU + ResBlocks
  + final LayerNorm) for a wider receptive field.
- FiLM (gamma, beta) modulation on top of the concat-then-mid path inside
  each LLLite module, zero-initialized to start from identity.
- Per-module depth embedding (zero-init bias) added to the shared cond_emb.
- Atomic target_layers specifiers, including a new mlp_fc1_pre target that
  injects LLLite into the GPT2FeedForward fc1.
- Optional ASPP tail in conditioning1, switchable via --lllite_use_aspp.

Weight key format (named, sd-scripts LoRA-style):
- Per-module keys are prefixed by the target Linear's full path, e.g.
  lllite_dit_blocks_0_self_attn_q_proj.down.weight, so the file alone
  identifies which Linear each tensor belongs to.
- Shared conditioning encoder is stored under lllite_conditioning1.*.
- depth_embeds is split per-module as {name}.depth_embed (cond_emb_dim,)
  for self-describability and easier per-block surgery later.
- Legacy formats (lllite_modules.{i}.* keys; v1/v2 in-development snapshots)
  are rejected on load with an explicit error.

Tests:
- networks/control_net_lllite_anima.py __main__ covers parse_target_layers,
  module count for all preset/atomic combinations, ASPP on/off, zero-init
  forward (3D and 5D), depth_embeds non-zero invariance, save/load
  round-trip with the new named keys, and legacy-format reject.
- tests/manual_test_anima_lllite_dryrun.py exercises build / apply_to /
  forward / backward (grads on LLLite only) / save+load round-trip on a
  stub DiT.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…ing in training function for ControlNet-LLLite

…ormat

The saved .safetensors uses a named-key format (per-module `{lllite_name}.*`
with `depth_embeds` split into per-module `{lllite_name}.depth_embed`), but
the doc described the internal state_dict layout (`lllite_modules.{i}.*` and
stacked `depth_embeds`). load_lllite_weights rejects the latter as legacy,
so the doc could mislead users into producing incompatible weight files.