first commit

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,2 @@
+.claude/
+drafts/

CLAUDE.md

@@ -0,0 +1,58 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Repository Overview
+
+A personal collection of ComfyUI workflows for AI video generation, video editing, upscaling, and image generation. Workflows are distributed as `.json` files that can be drag-and-dropped into ComfyUI. Each workflow lives in its own directory alongside a `README.md` and a `media/` folder with preview assets.
+
+## Directory Structure
+
+```
+<model-family>/<version>/<workflow-name>/
+    <workflow-name>.json    # The ComfyUI workflow
+    README.md               # Description, requirements, notes, changelog
+    media/                  # Preview images/videos (thumbnails, mp4s)
+
+posts/<date-slug>/
+    README.md               # Long-form making-of post
+    media/                  # Supporting media
+
+drafts/                     # Work-in-progress workflows (not published)
+```
+
+## Workflow Families
+
+- **ltx/** — LTX Video 2.x workflows (image-to-video, two-pass, audio support)
+- **wan/** — WAN 2.2 workflows (T2V, VACE, face detailing, upscaling, Lightning GGUF)
+- **hunyuanvideo/** — HunyuanVideo workflows (RF-inversion video editing)
+- **z-image/** — Image generation workflows (two-stage T2I with refinement)
+
+## README Convention for Workflows
+
+Each workflow README must follow this structure (keep sections in order, no extras):
+
+1. `# Title` — model name, type (Image-to-Video / Video-to-Video / etc.)
+2. One-paragraph description of what the workflow does
+3. `## Preview` — YouTube embed or local media link
+4. `## Requirements` — ComfyUI version, models with HuggingFace download links, custom nodes with GitHub links
+5. `## Notes` — key parameters, sampler settings, VRAM tips, optional nodes
+6. `## Changelog` — dated entries, newest first, format: `- \`YYYY-MM-DD\` — note`
+
+**No emojis** anywhere in README files. Use plain text markers if needed.
+
+## Adding a New Workflow
+
+1. Create the directory: `<model-family>/<version>/<workflow-slug>/`
+2. Export and place the `.json` from ComfyUI
+3. Create `README.md` following the convention above
+4. Create `media/` and add at minimum one preview thumbnail
+5. Add a row to the top-level `README.md` table under the correct model section (keep `| Workflow | Description | Type | Updated |` column order)
+
+## Updating the Main README
+
+The root `README.md` is the index. It contains one table per model family. Column order is fixed: `Workflow | Description | Type | Updated`. The `Type` column uses these values: `Video gen`, `Video edit`, `Upscaling`, `Image gen`.
+
+## Posts
+
+Long-form making-of articles go under `posts/<YYYY-MM-DD-slug>/README.md`. They are linked from the `## Posts` table in the root README. Posts reference workflows via relative paths (e.g. `../../ltx/2.3/i2v-two-pass/`).

README.md

@@ -1,3 +1,45 @@
----
-license: apache-2.0
----
+# ComfyUI Workflows
+
+A personal collection of ComfyUI workflows for AI video generation, upscaling, and video editing.
+Each workflow entry includes a description, example output, requirements, and usage notes.
+
+Drag-and-drop any `.json` file directly into ComfyUI to load the workflow.
+
+---
+
+## Posts
+
+| Post | Date |
+|---|---|
+| [MOMENTS — Making of](posts/2026-03-30-moments-short-film/) | 2026-03-30 |
+
+---
+
+## LTX Video 2.3
+
+| Workflow                                          | Description                                                                      | Type      | Updated    |
+| ------------------------------------------------- | -------------------------------------------------------------------------------- | --------- | ---------- |
+| [I2V Two-Pass](ltx/2.3/i2v-two-pass/)            | Two-pass I2V with start/end guide images, 720p→1080p upscale, and audio support  | Video gen | 2026-03-30 |
+
+## WAN 2.2
+
+| Workflow                                                  | Description                                                               | Type       | Updated    |
+| --------------------------------------------------------- | ------------------------------------------------------------------------- | ---------- | ---------- |
+| [T2V A14B](wan/2.2/t2v-a14b/)                             | Text-to-video with the WAN 2.2 A14B model, GGUF quantized                 | Video gen  | 2025-07-30 |
+| [Lightning GGUF](wan/2.2/lightning-gguf/)                 | Fast T2V with cached T5 encoding and Lightning-tuned steps                | Video gen  | 2025-08-09 |
+| [VACE](wan/2.2/vace/)                                     | Video-to-video with automatic subject segmentation and depth conditioning | Video edit | 2025-07-15 |
+| [VACE Endless Extension](wan/2.2/vace-endless-extension/) | Extend a video indefinitely using VACE looping                            | Video gen  | 2025-11-10 |
+| [Upscaling v1](wan/2.2/upscaling/)                        | Per-subject crop upscaling with VFI frame interpolation                   | Upscaling  | 2025-10-13 |
+| [Face Detailer](wan/2.2/face-detailer/)                   | SAM2-based segmentation and WAN 2.2 detail pass on existing video         | Video edit | 2026-03-29 |
+
+## HunyuanVideo
+
+| Workflow                                                    | Description                                                                  | Type       | Updated    |
+| ----------------------------------------------------------- | ---------------------------------------------------------------------------- | ---------- | ---------- |
+| [RF-Inversion](hunyuanvideo/rf-inversion/)                  | Video editing via RF-inversion — relight, restyle, or modify existing video  | Video edit | 2025-01-23 |
+
+## Z-Image
+
+| Workflow                        | Description                                             | Type      | Updated    |
+| ------------------------------- | ------------------------------------------------------- | --------- | ---------- |
+| [Two-Stage](z-image/two-stage/) | Two-pass T2I workflow with refinement and color grading | Image gen | 2026-03-29 |

hunyuanvideo/rf-inversion/README.md

@@ -0,0 +1,3 @@
+## Changelog
+
+- `2025-01-23` — Initial upload

hunyuanvideo/rf-inversion/hunyuanvideo-rf-inversion.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95431ecf28978a833268b76e51b5e7f88bdf5a2598ef2fc695087766af3d7836
+size 47453

ltx/2.3/i2v-two-pass/README.md

@@ -0,0 +1,47 @@
+# LTX Video 2.3 — I2V Two-Pass with Audio
+
+**Model:** LTX Video 2.3 — 22B Distilled (FP8)
+**Type:** Image-to-Video
+
+Two-pass Image-to-Video workflow using LTX Video 2.3 with audio generation support. Takes two guide images (start and end frame) and generates a video in a 720p first pass, then upscales to 1080p in a second pass. Audio is encoded via the LTX 2.3 audio VAE and composited into the final output.
+
+---
+
+## Preview
+
+[![Watch on YouTube — click to open](https://img.youtube.com/vi/5pMIWq4kwo8/maxresdefault.jpg)](https://www.youtube.com/watch?v=5pMIWq4kwo8)
+*Click to open on YouTube*
+
+---
+
+## Requirements
+
+- **ComfyUI:** recent stable build
+- **Model:** `ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors` — [Kijai/LTX2.3_comfy](https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/diffusion_models/ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors)
+- **Video VAE:** `LTX23_video_vae_bf16.safetensors` — [Kijai/LTX2.3_comfy](https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/vae/LTX23_video_vae_bf16.safetensors)
+- **Audio VAE:** `LTX23_audio_vae_bf16.safetensors` — [Kijai/LTX2.3_comfy](https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/vae/LTX23_audio_vae_bf16.safetensors)
+- **Preview VAE:** `taeltx2_3.safetensors` — [Kijai/LTX2.3_comfy](https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/vae/taeltx2_3.safetensors)
+- **Text encoder:** `gemma_3_12B_it_fp8_scaled.safetensors` — [Kijai/LTX2.3_comfy](https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/text_encoders)
+- **Text projection:** `ltx-2.3_text_projection_bf16.safetensors` — [Kijai/LTX2.3_comfy](https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/text_encoders/ltx-2.3_text_projection_bf16.safetensors)
+- **Custom nodes:**
+  - [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
+  - [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
+  - [rgthree-comfy](https://github.com/rgthree/rgthree-comfy)
+  - [ComfyUI_Fill-Nodes](https://github.com/filliptm/ComfyUI_Fill-Nodes)
+
+---
+
+## Notes
+
+- The workflow uses two guide images (start and end frame) via `LTXVAddGuideMulti`. Image 1 is placed at frame 0, Image 2 at frame 161 by default.
+- **Pass 1:** 1280x720, 241 frames @ 24fps, 8 steps, `lcm` sampler, `linear_quadratic` scheduler, CFG 1. NAG guidance is enabled (scale 11, rescale 0.25).
+- **Pass 2:** The output is upscaled to 1920x1080 with `nvidia_rtx_vsr` via `ImageResizeKJv2`, then re-encoded and sampled again with `euler_ancestral` sampler for 8 steps.
+- An audio file can be supplied via `VHS_LoadAudioUpload` — the audio is encoded with the LTX 2.3 audio VAE and blended into the latent space during both passes.
+- A LoRA (`ltx-2.3-22b-distilled-lora-dynamic`) is present in the graph but **disabled** — enable it if you want dynamic motion tuning.
+- `LTX2MemoryEfficientSageAttentionPatch`, `LTXVChunkFeedForward`, and `ModelPatchTorchSettings` are used for VRAM optimization. Requires SageAttention installed.
+
+---
+
+## Changelog
+
+- `2026-03-30` — Initial upload

ltx/2.3/i2v-two-pass/ltx2.3-i2v-two-pass.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60d153bde46947dd4b6888ad971d9a04c3df38a1619c7a9ed1bd4b3f06a53f2e
+size 96759

ltx/2.3/i2v-two-pass/media/preview.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b67918595c92b6826c32402148576e910af3065de7910f28412bee1a7d497101
+size 355345

posts/2026-03-30-moments-short-film/README.md

@@ -0,0 +1,91 @@
+# MOMENTS — Making of
+
+*2026-03-30*
+
+Lately I've been watching a lot of podcasts. Scientists, philosophers, that kind of thing. Neil deGrasse Tyson is a big favorite for me. At some point I thought — I want to make something like this. Just a short video, one idea, one voice talking directly to camera.
+
+Then the ArcaGidan challenge came out with the theme of Time, and that was the trigger. The first thing that popped into my head was that old saying — it's not the destination, it's the journey. I kept thinking about it. The journey isn't just the road... It's made of specific moments. The ones where things actually changed. Where you went one way instead of another, whether you chose it or not.
+
+And that's when a quote I'd known for years suddenly fit perfectly — Soren Kierkegaard: "Life can only be understood backwards; but it must be lived forwards." You're in the middle of something and you have no idea what it is yet. You only get it years later, when something random throws you right back there.
+
+Those things together became MOMENTS: a ~2-minute short film about the pivot points of a life.
+
+---
+
+## Preview
+
+[![Preview — MOMENTS teaser](media/cseti_moments_teaser_web_thumb.png)](media/cseti_moments_teaser_web.mp4)
+
+Watch the full short-film here: https://arcagidan.com/entry/06c37d91-060c-4a91-89e1-fae856888085
+
+If you like it, please vote — that's the best way to support this kind of work and help me keep making more opensource stuff: new workflows, fine-tuned models (full finetune, LoRA, IC-LoRA...), and so on.
+
+---
+
+## Building the Pipeline
+
+The basic flow was: write the script → write the prompts → generate a key frames → animate them → fix the face → upscale → add voice → add music → cut it together.
+
+### Narration: VibeVoice
+
+I didn't want something that obviously sounds like AI. I needed a voice that felt like someone actually sitting there and talking. I used [VibeVoice](../../) with a cloned voice based on P.J. Taylor from LibriVox ([reader 9165](https://librivox.org/reader/9165)). 
+
+### Key-frames: Two-Stage T2I → I2V
+
+So what I do is: first I generate one or more still image, get them looking exactly how I want it, then use those images for the videos. That's it basically.
+
+If I need key-frames for specific shots, I take the base image and manipulate it with an image-edit model like Qwen-Image-Edit or Nanobanana. 
+
+For the stills I used [Z-Image Two-Stage](../../z-image/two-stage/) — it does two passes, the second one at low denoise to clean up details.
+
+![Key-frames](media/keyframes-montage.jpg)
+
+### Animation: Image-to-Video
+
+I tried a lot of different workflows and settings for I2V. Still can't say I've figured it out — every project feels like I'm half-guessing. But here's one I put together that worked well for this film: [LTX 2.3 two-pass](../../ltx/2.3/i2v-two-pass/)
+
+The key-frames go in and the video prompt only describes motion, nothing visual. The visuals are already in the image.
+
+### Face Detailing: WAN 2.2
+
+LTX 2.3 has a lot of artifacts, especially on smaller details — faces, hands, fine textures. It's just how it is. F
+
+So I made a WAN 2.2 based detailer workflow for this. It segments the face - or anything else you want actually - with SAM2, crops it out, runs a WAN 2.2 pass on just that region, then puts it back into the original video. The rest of the clip stays untouched.
+
+The workflow is here: [WAN 2.2 Face Detailer](../../wan/2.2/face-detailer/)
+
+<table><tr>
+<td><a href="../../wan/2.2/face-detailer/media/preview-meeting.mp4"><img src="../../wan/2.2/face-detailer/media/preview-meeting_thumb.png" width="100%"></a></td>
+<td><a href="../../wan/2.2/face-detailer/media/preview-backwardlook.mp4"><img src="../../wan/2.2/face-detailer/media/preview-backwardlook_thumb.png" width="100%"></a></td>
+</tr></table>
+
+### Music: Suno
+
+The prompt I used:
+
+> Minimalist ambient bed for spoken word; slowly evolving piano clusters over warm synth swells and distant reversed textures. Very soft sidechained pads breathe around the voice, with subtle granular noise drifting in and out. Energy stays low and steady, occasionally blooming into gentle chords before receding, perfect for underscoring intimate storytelling, emotional.
+
+Suno did well with this.
+
+### Editing: Kdenlive
+
+Nothing fancy here. The script was written with the cuts already in mind, so putting it together in Kdenlive was mostly just mechanical assembly.
+
+### Tools at a glance
+
+| Role            | Tool / Workflow                                               |
+| --------------- | ------------------------------------------------------------- |
+| Music           | Suno                                                          |
+| TTS             | VibeVoice — cloned voice (P.J. Taylor / LibriVox reader 9165) |
+| Image/Video gen | ComfyUI                                                       |
+| Image-to-Video  | [LTX 2.3 two-pass](../../ltx/2.3/i2v-two-pass/)               |
+| Text-to-Image   | [Z-Image Two-Stage](../../z-image/two-stage/)                 |
+| Image editing   | Qwen-Image / Nanobanana                                       |
+| Face detailing  | [WAN 2.2 Face Detailer](../../wan/2.2/face-detailer/)         |
+| Video editing   | Kdenlive                                                      |
+
+---
+
+## Changelog
+
+- `2026-03-30` — Published

posts/2026-03-30-moments-short-film/media/cseti_moments_teaser_web.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e32a555dcae4dbd5499e4d0807ba6dcdbf77bce4a3e93fdb3c62df0ec93db17d
+size 9213227

wan/2.2/face-detailer/README.md

@@ -0,0 +1,64 @@
+# WAN 2.2 Face Detailer
+
+**Model:** WAN 2.2 T2V A14B (fp8 scaled)
+**Type:** Video-to-Video
+
+Face enhancement pipeline for WAN 2.2 video output. Loads an existing video, detects and segments faces using SAM2, crops each face region, runs a focused WAN 2.2 generation pass to add detail, then composites the result back into the original frame.
+
+---
+
+## Preview
+
+[![Preview — meeting scene](media/preview-meeting_thumb.png)](media/preview-meeting.mp4)
+
+[![Preview — backward look scene](media/preview-backwardlook_thumb.png)](media/preview-backwardlook.mp4)
+
+---
+
+## Requirements
+
+- **ComfyUI:** recent stable build
+
+### Models
+
+- **Main model:** `Wan2_2-T2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors` —
+  [Kijai/WanVideo_comfy_fp8_scaled](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/resolve/main/T2V/Wan2_2-T2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors)
+
+- **VACE module:** `Wan2_2_Fun_VACE_module_A14B_LOW_fp8_e4m3fn_scaled_KJ.safetensors` —
+  [Kijai/WanVideo_comfy_fp8_scaled](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/resolve/main/VACE/Wan2_2_Fun_VACE_module_A14B_LOW_fp8_e4m3fn_scaled_KJ.safetensors)
+
+- **VAE:** `Wan2_1_VAE_bf16.safetensors` —
+  [Kijai/WanVideo_comfy](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan2_1_VAE_bf16.safetensors)
+
+- **Text encoder:** `umt5-xxl-enc-bf16.safetensors` —
+  [Kijai/WanVideo_comfy](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/umt5-xxl-enc-bf16.safetensors)
+
+- **LoRA (step distill):** `lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16.safetensors` —
+  [Kijai/WanVideo_comfy](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16.safetensors)
+
+- **LoRA (reward):** `Wan2.2-Fun-A14B-InP-low-noise-HPS2.1.safetensors` — example LoRA, replace with your own if needed —
+  [alibaba-pai/Wan2.2-Fun-Reward-LoRAs](https://huggingface.co/alibaba-pai/Wan2.2-Fun-Reward-LoRAs/resolve/main/Wan2.2-Fun-A14B-InP-low-noise-HPS2.1.safetensors)
+
+- **Segmentation:** `sam2.1_hiera_small.safetensors` —
+  [Kijai/sam2-safetensors](https://huggingface.co/Kijai/sam2-safetensors/resolve/main/sam2.1_hiera_small.safetensors)
+
+### Custom nodes
+
+- [ComfyUI-WanVideoWrapper](https://github.com/kijai/ComfyUI-WanVideoWrapper)
+- [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
+- [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
+- [ComfyUI-segment-anything-2](https://github.com/kijai/ComfyUI-segment-anything-2)
+- [rgthree-comfy](https://github.com/rgthree/rgthree-comfy)
+- [ComfyUI-Florence2](https://github.com/kijai/ComfyUI-Florence2) — delete these nodes if you don't want to use florence2 for segmentation.
+---
+
+## Notes
+
+- The VACE module is optional — disconnect it to run without VACE.
+- SAM2 is used for face segmentation. The face bounding box index can be adjusted directly in the workflow if the wrong face is selected.
+
+---
+
+## Changelog
+
+- `2026-03-29` — Initial upload

wan/2.2/face-detailer/media/preview-backwardlook.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdc913059a210cddd43ee12ad80fa2c73aad205e30351d778ea08f8a591262c0
+size 734426

wan/2.2/face-detailer/media/preview-meeting.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0301b94b775cf3dc5c943c8f56b6b87e5a1bc2a0db9efae6db802ba96fc04b5b
+size 1082711

wan/2.2/face-detailer/wan2.2-face-detailer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a51bc897a65aadac4623d21dc9e85d5c0b044781d194cda5645711714f3fb63d
+size 262508

wan/2.2/lightning-gguf/README.md

@@ -0,0 +1,3 @@
+## Changelog
+
+- `2025-08-09` — Initial upload

wan/2.2/lightning-gguf/media/preview.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f47907d5ee9ca478ba68a5a3d25459b7bea17a138318a08cb2dd73d8b570dfe
+size 1226780

wan/2.2/lightning-gguf/wan2.2-lightning-gguf.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e605ebccb8f061cf20f3858826e3a57f16c8413b50762cd98fdea62768a7b7a
+size 42711

wan/2.2/t2v-a14b/README.md

@@ -0,0 +1,3 @@
+## Changelog
+
+- `2025-07-30` — Initial upload

wan/2.2/t2v-a14b/media/preview.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3310a49cd6a46c315965f02994aa6812622b7f09d90396a5ab1bfc819d0f589
+size 2459097

wan/2.2/t2v-a14b/wan2.2-t2v-a14b.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e39c02428dd6554640ab0dae22f25ef366a7829c4d42daf7af0b1b84ee515dc8
+size 38151

wan/2.2/upscaling/README.md

@@ -0,0 +1,57 @@
+# WAN 2.2 Upscaling v1
+
+**Model:** WAN 2.2 A14B LowNoise (GGUF Q4_K_M)
+**Type:** Upscaling (Video-to-Video)
+
+Per-subject crop upscaling pipeline for WAN 2.2 video output.
+Subjects and faces are detected automatically via SAM2 segmentation and Florence2 face detection,
+then cropped, upscaled with pixel-based models, refined with WAN video-to-video, and composited back.
+RIFE VFI is used for frame interpolation on the final output.
+
+---
+
+## Preview
+
+![Preview v2 (looping)](media/preview-v2-looping.png)
+
+---
+
+## Requirements
+
+- **ComfyUI:** recent stable build
+- **Model:** `wan2.2/Wan2.2-T2V-A14B-LowNoise-Q4_K_M.gguf` — [download](https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main) (place in `ComfyUI/models/unet/`)
+- **VAE:** `wan/Wan2_1_VAE_bf16.safetensors` — [download](https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan2_1_VAE_bf16.safetensors) (place in `ComfyUI/models/VAE/`)
+- **Text encoder:** `umt5-xxl-enc-bf16.safetensors` — [download](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/text_encoders) (place in `ComfyUI/models/text_encoders/`)
+- **LoRA (speed):** `Wan/lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16.safetensors` — [download](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Lightx2v)
+- **LoRA (quality, example):** `Wan/wan2.2_fun-reward/Wan2.2-Fun-A14B-InP-low-noise-HPS2.1.safetensors` — reward LoRA for quality improvement; replace with any WAN 14B compatible LoRA or remove
+- **SAM2 model:** `sam2.1_hiera_base_plus.safetensors` — auto-downloaded by `DownloadAndLoadSAM2Model`
+- **RIFE model:** `rife49.pth` — auto-downloaded by `RIFE VFI`
+- **Upscale models** (place in `ComfyUI/models/upscale_models/`):
+  - `4x-ClearRealityV1.pth` — [download](https://openmodeldb.info/models/4x-ClearRealityV1) — fast photo-realistic upscaler
+  - `phhofm/1xDeJPG_realplksr_otf.pth` — [download](https://openmodeldb.info/models/1x-DeJPG-realplksr-otf) — JPEG artifact remover / enhancer
+  - `1xSkinContrast-HighAlternative-SuperUltraCompact.pth` — [download](https://openmodeldb.info/models/1x-SkinContrast-HighAlternative-SuperUltraCompact) — skin detail enhancer
+  - `DF2K_JPEG.pth` — [download](https://openmodeldb.info/models/4x-realsr-df2k-jpeg) — slower but higher quality upscaler
+- **Custom nodes:**
+  - [ComfyUI-WanVideoWrapper](https://github.com/kijai/ComfyUI-WanVideoWrapper)
+  - [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
+  - [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
+  - [ComfyUI-Florence2](https://github.com/kijai/ComfyUI-Florence2)
+  - [ComfyUI-segment-anything-2](https://github.com/kijai/ComfyUI-segment-anything-2)
+  - [ComfyUI-Frame-Interpolation](https://github.com/Fannovel16/ComfyUI-Frame-Interpolation)
+  - [rgthree-comfy](https://github.com/rgthree/rgthree-comfy)
+
+---
+
+## Notes
+
+- **Face index:** After running Florence2, each detected face gets a numbered ID visible above the bounding box in the preview. Set the `Index` field in the face crop node to the ID of the face you want to upscale.
+- **crop_size_mult:** Controls the crop area around the face. Use a higher value for faces that are large on screen, lower for small faces. Follow the preview output to calibrate.
+- **VRAM:** For 1080p output on under 16 GB VRAM, reduce `context_frames` to 33 or 41 in `WanVideoContextOptions`.
+- **WAN denoise strength:** Upscale passes use 0.3 denoise — low enough to preserve structure, high enough for texture refinement.
+- Input resolution: 832x480, 81 frames. The pipeline outputs at a higher resolution after upscaling.
+
+---
+
+## Changelog
+
+- `2025-10-13` — Initial upload

wan/2.2/upscaling/wan2.2-upscaling-v1.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8045ccdb02922a699d81e78c54a9aeb9d52a361f3469ca1ae29b3cdae2ad54ca
+size 265934

wan/2.2/vace-endless-extension/README.md

@@ -0,0 +1,3 @@
+## Changelog
+
+- `2025-11-10` — Initial upload

wan/2.2/vace-endless-extension/wan2.2-vace-endless-extension.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0759d7a20d0b7e50eeaccaad5e1eb1df840333cc0a2192f6a2dbd0f15a11bb97
+size 231460

wan/2.2/vace/README.md

@@ -0,0 +1,3 @@
+## Changelog
+
+- `2025-07-15` — Initial upload

wan/2.2/vace/media/preview.webm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c5a387ccf4b4aadfb339f58dd01b1be31e327ffe66129365e6be3e84ac449c1
+size 1881158

wan/2.2/vace/wan2.2-vace.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:806f7513c18449be7dacfa15cfc999e5f0317f12a5de171dc12dfc81e6429359
+size 78894

z-image/two-stage/README.md

@@ -0,0 +1,41 @@
+# Z-Image Two-Stage
+
+**Model:** Z-Image (z_image_turbo) + MoodyMix DPO refinement
+**Type:** Text-to-Image
+
+Two-pass text-to-image workflow using Comfy-Org's Z-Image model. The first pass generates a full-resolution base image with the BF16 base model; the second pass applies a DPO-tuned refinement model at low denoise strength to improve detail and style consistency. A GLSL-based post-processing subgraph handles HSL color grading.
+
+---
+
+## Preview
+
+![Preview](media/preview.png)
+
+---
+
+## Requirements
+
+- **ComfyUI:** recent stable build (subgraph support required, frontend >= 1.43)
+- **Model (Stage 1):** `z-image/z_image_bf16.safetensors` — [download](https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/diffusion_models/z_image_turbo_bf16.safetensors)
+- **Model (Stage 2):** `z-image/moodyMix_zitV10DPO.safetensors` — [CivitAI](https://civitai.com/models/620406) (verify V10 DPO version)
+- **VAE:** `Flux/ultraflux.safetensors` — [CivitAI](https://civitai.com/models/2231253) — alternatively `ae.safetensors` from the official split ([download](https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors))
+- **Text encoder:** `qwen_3_4b.safetensors` (lumina2) — [download](https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors)
+- **Custom nodes:**
+  - [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) — `DiffusionModelLoaderKJ`, `ImageResizeKJv2`, `Film Grain`, and others
+  - [rgthree-comfy](https://github.com/rgthree/rgthree-comfy) — `Seed (rgthree)`, `Image Comparer (rgthree)`
+
+---
+
+## Notes
+
+- **Stage 1:** 20 steps, CFG 4, sampler `res_multistep`, scheduler `simple`, full denoise
+- **Stage 2:** 2 steps, CFG 1, sampler `res_multistep`, denoise 0.36 — intended as a light refinement pass
+- Base latent resolution: 1920×1088; the second pass includes a 2× upscale step
+- The color grading subgraph applies global HSL adjustments (default: saturation -10, all else neutral); it can be bypassed or tuned per shot
+- The workflow uses ComfyUI subgraphs — load it in a recent frontend build that supports the subgraph renderer (`LG`)
+
+---
+
+## Changelog
+
+- `2026-03-29` — Initial upload

z-image/two-stage/z-image-two-stage.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe19f97eca330bc9c4f22a1b4a924ebe34c378d4aae14e39ae70885f8d334edd
+size 103598