nodes/havvk-ComfyUI_AIIA
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
1,081 Commits 1 Branch 0 Tags
main
Commit Graph
325 Commits
Author SHA1 Message Date
Hawk Lee f73661f655 feat: release version 1.15.0 with IndexTTS-2 inline emotion tags and fixes 2026-02-20 01:17:02 +08:00
Hawk Lee 515d749177 fix: remove int16 clipping and use float32 output in infer_v2.py 2026-02-19 17:09:23 +08:00
Hawk Lee f7e96d37ce fix: 用 transformers 替代 modelscope 导入，从根源消除副作用 
- infer_v2.py: from modelscope → from transformers import AutoModelForCausalLM
  模型已在本地，不需要 modelscope 的 hub 下载包装
  这样 modelscope 根本不会被 import，消除所有 monkey-patching 副作用
- aiia_vibevoice_nodes.py: 还原 device_map=auto（此文件无需修改）
 2026-02-19 12:51:36 +08:00
Hawk Lee 8b3b73c2ed fix: QwenEmotion device_map=auto 与 modelscope 不兼容导致 IndexError 
- 移除 device_map='auto'，改为手动 .to(device)
- 与 IndexTTS2 中其他模型(GPT/BigVGAN/CamPPlus)加载方式一致
- 修复 use_cuda_kernel=False 时的加载失败
 2026-02-19 12:39:41 +08:00
Hawk Lee 1cb441eaa9 fix: 修复 IndexTTS 导入导致 NeMo 分段节点卡死的问题 
- 根因: indextts 的 infer_v2.py 顶层导入 modelscope，会全局 monkey-patch
  transformers.PreTrainedModel.from_pretrained，导致 NeMo 模型加载卡死
- 修复: 在 aiia_indextts_nodes.py 中保存/恢复 from_pretrained 原始引用
- 清理 libs/ 目录：删除各依赖的文档、示例、测试等非运行时文件
- 新增 libs/index-tts (仅保留 indextts/ 核心包和 LICENSE)
 2026-02-19 12:05:07 +08:00
Hawk Lee f95f2a9832 v1.14.0: Fix CosyVoice transformers 4.57+ compatibility & emotion segment crossfade 
- Qwen2Encoder: bypass incompatible Qwen2Model.forward() in transformers ≥4.53
  via manual layer iteration + SDPA-compatible mask + POST-norm output
- Add cosine crossfade (50ms) between multi-emotion TTS segments
- Update README compatibility table to reflect full transformers support
 2026-02-17 15:49:12 +08:00
Hawk Lee e21a5c9eba fix: restore file permissions to 755 (reverts eff537d mode change) 
The eff537d commit changed all files from 755 to 644.
Restores original executable permissions.
 2026-02-16 12:15:32 +08:00
Hawk Lee 68d71bb728 fix: complete EchoMimicV3 dist stubs for single-GPU inference 
Add all missing distributed utility stubs:
- get_sequence_parallel_rank (returns 0)
- get_sequence_parallel_world_size (returns 1)
- get_sp_group (returns None)
- set_multi_gpus_devices (no-op)
- xFuserLongContextAttention (stub class)
- wan_xfuser.usp_attn_forward (stub)
 2026-02-16 01:37:15 +08:00
Hawk Lee e6e17aec81 fix: add proxy_url input + EchoMimicV3 dist stub 
- Replace ~/run.sh reading with proxy_url node input parameter
- Proxy priority: node param > HTTPS_PROXY env var
- Add no-op parallel_magvit_vae stub for single-GPU inference
- Remove environment-specific ~/run.sh dependency
 2026-02-16 01:34:21 +08:00
Hawk Lee eff537d9dd feat: add ASR, Podcast Splitter, and Podcast Stitcher nodes for anti-leakage pipeline 
- Add AIIA_ASR node with FunASR (paraformer-zh + SenseVoiceSmall)
- Add AIIA_Podcast_Splitter for per-speaker text splitting
- Add AIIA_Podcast_Stitcher with 3-tier fuzzy alignment (exact, Levenshtein, fallback)
- Register 5 previously missing modules in __init__.py
- Update README with installation guide and node documentation
 2026-02-15 12:14:45 +08:00
Hawk Lee de9eefcd64 v1.10.13: Fix Ditto Sampler interruption issue 2026-02-02 14:20:52 +08:00
Hawk Lee 26b6a1d368 v1.9.700: Active Trajectory Planner & Vitality Boost 2026-01-26 21:05:46 +08:00
Hawk Lee ef8aa8cdf5 Fix(Ditto): v1.9.510 Hard Lock silence + Fix Noise Leak. Disable Stitcher Buffer Clear. 2026-01-26 19:16:38 +08:00
Hawk Lee 641c03bb38 feat(Ditto): v1.9.505 Hybrid Overwrite - Lock Head to Ref, Release Expressions 2026-01-26 17:40:16 +08:00
Hawk Lee ba9e8d8f83 feat(Ditto): v1.9.504 Direct Output Overwrite - Bypass AI during silence to guarantee zero drift 2026-01-26 17:36:55 +08:00
Hawk Lee a4abb7b248 fix(Ditto): v1.9.503 Ramped Gravity + Faster Return Speed to fix shakes and drift accumulation 2026-01-26 17:26:40 +08:00
Hawk Lee 8e1313af01 fix(Ditto): v1.9.502 FINAL - Delete Zombie Code, Boost Idle Gravity, Reduce Pressure 2026-01-26 17:15:15 +08:00
Hawk Lee 90d3d07f89 fix(Ditto): v1.9.501 force new version string to debunk caching issues 2026-01-26 17:04:37 +08:00
Hawk Lee 27e3fedc90 feat(Ditto): v1.9.500 Procedural Idle - switch to deterministic interpolation during silence to guarantee stability 2026-01-26 16:44:05 +08:00
Hawk Lee faf3b001b9 fix(Ditto): v1.9.402 restore warp decay during speech and reduce idle pressure (0.8->0.4) to fix reference lock 2026-01-26 16:35:56 +08:00
Hawk Lee b88f435754 fix(Ditto): increase silence gravity to 0.20 to prevent AI model internal drift/droop 2026-01-26 16:21:29 +08:00
Hawk Lee ea3729b171 fix(Ditto): DELETE zombie anchor code causing infinite drift feedback loop 2026-01-26 16:14:36 +08:00
Hawk Lee c462f2606b fix(Ditto): bump version to v1.9.401 and clean up debug crash 2026-01-26 15:59:53 +08:00
Hawk Lee 4e289fb581 debug(Ditto): inject runtime error (nuclear option) to force crash and reveal load path 2026-01-26 15:56:01 +08:00
Hawk Lee acb31d699d fix(Ditto): add granular debug logs for drift physics diagnosis 2026-01-26 15:49:05 +08:00
Hawk Lee dddef52cb6 fix(Ditto): add missing import os for debug logger 2026-01-26 15:11:51 +08:00
Hawk Lee 644267d956 fix(Ditto): add file path debug print to diagnose deployment location 2026-01-26 15:07:49 +08:00
Hawk Lee 87196eadc4 fix(Ditto): add startup log to verify new physics engine (Mean-Reverting) is loaded 2026-01-26 14:59:42 +08:00
Hawk Lee cec2bfebb6 fix(Ditto): replace runaway momentum physics with stable mean-reverting (OU) process for drift 2026-01-26 14:50:05 +08:00
Hawk Lee d7514af23b fix(Ditto): disable current_push and recovery bias to stop runaway head droop (drift accumulation) 2026-01-26 14:41:10 +08:00
Hawk Lee 07dc6d4658 fix(Ditto): aggressively decay momentum(0.5) and position(0.9) during silence to stop runaway drift 2026-01-26 14:33:22 +08:00
Hawk Lee cdcb9b645a fix(Ditto): move pressure reset to START of frame logic to prevent onset glitch 2026-01-26 14:09:48 +08:00
Hawk Lee a5fc09a91a fix(Ditto): syntax error in diagnostic logging block 2026-01-26 13:59:52 +08:00
Hawk Lee 928309b099 fix(Ditto): decay brownian drift during silence (>25f) to align anchor with reference photo 2026-01-26 13:52:11 +08:00
Hawk Lee 41621e8fc6 fix(Ditto): force instant pressure release at onset to rely 100% on Warp Offset (prevent anchor jump) 2026-01-26 13:16:34 +08:00
Hawk Lee 322e06502f fix: use self.motion_stitch instead of undefined local var 2026-01-26 11:58:05 +08:00
Hawk Lee 724d4fc029 fix: indentation error in stream_pipeline_offline.py 2026-01-26 11:53:49 +08:00
Hawk Lee b2da958dc6 fix(Ditto): clear StitchNetwork buffer on speech onset to prevent smoothing glitches 2026-01-26 11:33:17 +08:00
Hawk Lee 30c3d9d581 fix(Ditto): active return to ref (press=0.8, step=0.01) and disabled lookahead for immediate drift 2026-01-26 11:20:08 +08:00
Hawk Lee aec78ce5a3 fix(Ditto): tuned idle return params (press=0.5, step=0.004, decay=0.96) for balanced responsiveness 2026-01-26 11:08:26 +08:00
Hawk Lee 95f97a8482 fix(Ditto): removed aggressive ceiling; softened hysteresis push (0.0045->0.001) but triggered earlier (-10.0) for safe connection 2026-01-26 10:46:25 +08:00
Hawk Lee c69b6d8e2a fix(Ditto): corrected pitch limit logic (< -5.0) and reduced sway freq/momentum for stability 2026-01-26 02:09:42 +08:00
Hawk Lee 2357d229bf fix(Ditto): reduced noise/jitter and added pitch ceiling to fix stability and mouth opening 2026-01-26 02:01:21 +08:00
Hawk Lee b8ae3f2e9e fix(Ditto): ultra-smooth warp decay (0.98) and pressure (0.002) to prevent sliding 2026-01-26 01:47:03 +08:00
Hawk Lee b3f4eb9cf3 fix(Ditto): relaxed idle pressure (0.6->0.3) and recovery threshold (-8->-15) to prevent unwanted twisting 2026-01-26 01:23:58 +08:00
Hawk Lee e5eef09e02 fix: correct indentation in stream_pipeline_offline.py 2026-01-26 01:06:24 +08:00
Hawk Lee 8a53e2f4f7 fix(Ditto): smooth pressure release on speech onset to prevent teleport 2026-01-26 00:58:18 +08:00
Hawk Lee c8eb5a327e fix(Ditto): disable forced physical reset on speech onset 2026-01-26 00:46:48 +08:00
Hawk Lee 7a99e9afaf Ditto Surgical Update (v1.9.400): Rational Reset. Fixed Speech Onset Snap via Instant Pressure Release, 0:202 Z-Axis Alignment, and Latent Gravity Suppression. 2026-01-25 19:32:18 +08:00
Hawk Lee 9f618dacc0 v1.9.318: Emergency Legacy Restoration (Rollback to v1.9.224 state for stability) 2026-01-25 19:12:39 +08:00