Add creative/comfyui v5.1.0 (official bundled) and media/ltx-video-pipeline v2.0.0 (custom LTX pipeline)

2026-07-22 16:17:52 -05:00
parent 7e81fdac31
commit d94132f4ca
60 changed files with 10421 additions and 0 deletions
@@ -0,0 +1,86 @@
+# Video Generation Research — Qdrant Quick-Start
+
+**Canonical data store**: Qdrant `comfyui_decisions` @ `http://10.0.0.22:6333` (64+ points)
+**Embedding**: `snowflake-arctic-embed2:latest` (1024-dim, Cosine)
+**Human companion**: `/home/n8n/workspace/comfy/research.md` (search topics + URLs only, no specs)
+
+---
+
+## What Lives Where
+
+| Store | Content | How to Query |
+|-------|---------|-------------|
+| **Qdrant** `comfyui_decisions` | All specs, VRAM numbers, node names, workarounds, paper abstracts, trend analysis | Vector search via Ollama embedding + POST to 10.0.0.22:6333 |
+| **Qdrant** `comfyui_kb` | General knowledge (3884 points, pre-existing) | Same method |
+| **research.md** | Human-readable search topics, URL quick-reference, comparison tables | grep / raw read |
+| **This file** | Quick-start pointers to the above | Read once, then query Qdrant |
+
+---
+
+## Qdrant Query Pattern
+
+```python
+import requests, json
+
+# 1. Get embedding
+emb_resp = requests.post("http://localhost:11434/api/embed", json={
+    "model": "snowflake-arctic-embed2:latest",
+    "input": "YOUR SEARCH TOPIC"
+})
+vector = emb_resp.json()["embeddings"][0]
+
+# 2. Search Qdrant
+search_resp = requests.post(
+    "http://10.0.0.22:6333/collections/comfyui_decisions/points/search",
+    json={"vector": vector, "limit": 5, "with_payload": True}
+)
+results = search_resp.json()["result"]["points"]
+```
+
+---
+
+## Example Search Topics
+
+Use these as `input` strings above (from `research.md`):
+
+- "LTX 2.3 specs, frames, VRAM, audio latent"
+- "HunyuanVideo frame limits block swapping FP8"
+- "Wan 2.2 low VRAM sound-to-video FreeLong"
+- "Google Veo 3 native audio cloud"
+- "Kling 3.0 Omni visual identity vocal tone"
+- "ID-LoRA face identity lock LTX"
+- "FreeLong spectral blending Wan motion reversal"
+- "Segment chaining last frame to first frame"
+- "LivePortrait portrait animation"
+- "MuseTalk LatentSync lip-sync"
+- "RIFE frame interpolation"
+- "MultiTalk NeurIPS multi-person conversation"
+- "StreamingT2V CVPR long video streaming"
+
+---
+
+## Quick Model Comparison (from Qdrant)
+
+| Tool | Native Audio | Max Frames | VRAM | Best For |
+|------|-------------|------------|------|----------|
+| LTX 2.3 | YES | 361 | 12-24 GB | Local, audio-synced |
+| HunyuanVideo 1.5 | NO | 129 | 20-24 GB | Best open-source quality |
+| Wan 2.2 | YES (S2V) | 81 / 1025 | 5-16 GB | Low VRAM, sound-driven |
+| Veo 3 | YES | API | Cloud | Highest fidelity |
+| Kling 3.0 Omni | YES | API | Cloud | Character + voice |
+
+---
+
+## Key URLs (Quick Reference)
+
+- LTX: https://huggingface.co/Lightricks
+- Hunyuan: https://github.com/Tencent-Hunyuan/HunyuanVideo
+- Wan wrapper: https://github.com/kijai/ComfyUI-WanVideoWrapper
+- Veo: https://deepmind.google/models/veo/
+- Kling: https://kling.ai/
+- LivePortrait: https://github.com/kijai/ComfyUI-LivePortraitKJ
+- VHS: https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite
+
+---
+
+*All detailed data is in Qdrant. This file is a navigation aid only.*