━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎬 **LTX-2.3 Image-to-Video Workflow**
QwenVL Auto-Prompt · No Drift · ComfyUI
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Pure LTX 2.3 22B image-to-video pipeline for ComfyUI. Drop an image, get professional motion. QwenVL vision model automatically analyzes your input image and generates a motion-aware prompt—no manual description needed. The workflow enforces locked static camera (anti-drift), scales dynamically to any input resolution, and upscales output to broadcast-quality 1920×1088 at 24 FPS. Production-ready for stock footage, ambient loops, and commercial video generation.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✨ **Features**
✅ QwenVL Auto Motion Director — Vision model reads input image → auto-generates motion prompt with camera lock and object tracking hints
✅ Locked Static Camera — Zero pan, zoom, or drift; all motion in-frame only
✅ Pure LTX 2.3 22B — No LoRA needed; GGUF quantization for 16GB VRAM
✅ Dynamic Pixel Scaling — Auto-scales any input size to optimal 0.52MP for 8-step inference
✅ Dual-Stage Upscale — 960×544 base → 2× spatial upscaler → 1920×1088 output
✅ Audio + Video VAE — Multi-modal encoding; ready for synced audio pipelines
✅ 24 FPS Native — Smooth playback; 168 frames per generation
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📦 **Required Models** (6 files, ~32 GB)
• ltx-2.3-22b-distilled-Q4_K_M.gguf (17.8 GB) — Main UNet diffusion model (GGUF Q4 quantized)
• gemma_3_12B_it_fp4_mixed.safetensors (9.45 GB) — Text encoder for LTX prompt understanding
• ltx-2.3_text_projection_bf16.safetensors (2.31 GB) — Text-to-latent projection layer
• LTX23_video_vae_bf16.safetensors (1.45 GB) — Video VAE codec (encode/decode video frames)
• LTX23_audio_vae_bf16.safetensors (365 MB) — Audio VAE codec (dual-modal support)
• ltx-2.3-spatial-upscaler-x2-1.1.safetensors (996 MB) — 2× spatial upscaler for final quality pass
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⬇️ **Download Links** (verified HuggingFace)
📁 **ComfyUI/models/unet/**
• LTX-2.3-22B-distilled-1.1-Q4_K_M.gguf (17.8 GB) — https://huggingface.co/QuantStack/LTX-2.3-GGUF
📁 **ComfyUI/models/text_encoders/**
• gemma_3_12B_it_fp4_mixed.safetensors (9.45 GB) — https://huggingface.co/Comfy-Org/ltx-2
• ltx-2.3_text_projection_bf16.safetensors (2.31 GB) — https://huggingface.co/Kijai/LTX2.3_comfy
📁 **ComfyUI/models/vae/**
• LTX23_video_vae_bf16.safetensors (1.45 GB) — https://huggingface.co/Kijai/LTX2.3_comfy
• LTX23_audio_vae_bf16.safetensors (365 MB) — https://huggingface.co/Kijai/LTX2.3_comfy
📁 **ComfyUI/models/upscale_models/**
• ltx-2.3-spatial-upscaler-x2-1.1.safetensors (996 MB) — https://huggingface.co/Lightricks/LTX-2.3
⚠️ *VAE files are NOT in the official Lightricks repo — get them from Kijai/LTX2.3_comfy. Gemma fp4 encoder hosted by Comfy-Org. Filenames use v1.1 (current stable hotfix release).*
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🧩 **Required Custom Nodes**
• LTXV — Lightricks LTX-Video extension (sampling, encoding, projection)
• AILab_QwenVL_Advanced — QwenVL vision model integration for image-to-text
• ComfyUI-GGUF — UnetLoaderGGUF for quantized model loading
• VideoHelperSuite — VHS_VideoCombine, frame batching, video output export
• rgthree-comfy — Fast Groups Bypasser (optional; used for workflow flexibility)
• ImageIterator — Batch image loader for multi-image workflows
• ImageScaleToTotalPixels — Dynamic resolution scaling to pixel budget
• GetImageSize+ — Image dimension detection for auto-scaling pipeline
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🚀 **How to Use**
1. Place your input image(s) in the ComfyUI ./input directory
2. Load this workflow into ComfyUI
3. (Optional) Review the auto-generated motion prompt in the QwenVL output text node
4. Queue and generate
5. Output video saved via VHS to ./output directory
The entire motion prompt generation and scaling pipeline runs automatically—queue once, get your result.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚙️ **Settings & Parameters**
• FPS — 24 (Standard frame rate; 168 total frames per generation)
• Pixel Budget — 0.52 MP (Optimal for 8-step sampling on 16GB VRAM)
• Sampler — er_sde (Low-drift SDE solver for stable motion)
• Base Steps — 8 (Main diffusion sampling passes)
• Refine Steps — 3 (Quality refinement after upscale)
• CFG Scale — 1.0 (Classifier-free guidance; 1.0 = no guidance, stable output)
• Output Resolution — 1920×1088 (After 2× spatial upscale)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
💡 **Performance Tips**
• Batch Multiple Images — Queue 5–10 images in one session to amortize model load time
• Input Image Quality — Sharp, well-lit images yield sharper motion; low-contrast images may produce soft motion
• Motion Prompt Tuning — Edit the QwenVL text output node before queuing if you want specific motion direction (e.g., remove camera keywords to force static)
• Speed vs. Quality — The dual-stage upscale adds ~20 seconds per clip. Bypass the Spatial Upscaler node if speed is critical (output at 960×544)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📝 **Notes & AI Disclosure**
• AI-Generated Content — All example outputs are AI-generated by LTX 2.3. Suitable for stock footage, ambient loops, and creative projects.
• Model Downloads — See the "Download Links" section above for exact HuggingFace repos and target folders.
• Hardware Tested — RTX 5080 16GB VRAM; CUDA compute 9.2+
• VRAM Usage — ~14 GB peak during sampling; requires fast SSD for frame buffering
• No Commercial Guarantees — Use at your own discretion. Respect local AI disclosure laws when publishing outputs.
Enjoy clean, drift-free motion generation. Questions? Test the workflow locally first—Civitai comments section is for feedback, not troubleshooting.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔗 **Also check out**
New: **[SeedVR2 Batch Upscaler](https://civarchive.com/models/2750373/seedvr2-batch-upscaler-sleep-on-it-wake-up-4k?modelVersionId=3094090)** — Sleep On It, Wake Up 4K. Drop a whole folder of stills, walk away, come back to 4K. Great for upscaling your generations.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⭐ **Found this useful?**
• Like if it saved you time
• Comment your results — I read every one
• Follow for new ComfyUI workflows, all tested on 16 GB VRAM
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚖️ **Model Attribution & Licensing**
**LTX-Video 2.3** (Lightricks)
• License: LTX-2 Community License — https://huggingface.co/Lightricks/LTX-2.3/blob/main/LICENSE
• Free for commercial use by entities under $10M USD annual revenue
• AI-generated content disclosure required
**Gemma 3 12B IT** (Google DeepMind)
• License: Gemma Terms of Use — https://ai.google.dev/gemma/terms
• Subject to Google's Prohibited Use Policy
**Custom Nodes**
• LTXV (Lightricks), VideoHelperSuite (MIT), AILab QwenVL, rgthree-comfy (MIT), ComfyUI-GGUF
All example outputs are AI-generated. This workflow (JSON configuration) is shared as original work; model weights must be downloaded separately from the official sources above.
Description
Initial release. Pure LTX 2.3 22B i2v, QwenVL auto-prompt, locked camera, dual-stage upscale to 1920x1088.
FAQ
Comments (8)
Can't find the correct ImageIterator Node. Can you link it
The "ImageIterator" node comes from the ComfyUI_Image_Anything pack (not a standalone node).
📦 Repo: https://github.com/ComfyUI-Kelin/ComfyUI_Image_Anything
Install:
1. ComfyUI Manager → Install via Git URL → paste the link above
(or: cd ComfyUI/custom_nodes && git clone https://github.com/ComfyUI-Kelin/ComfyUI_Image_Anything.git)
2. Restart ComfyUI
3. The node shows up as "Image Iterator" in the menu
Let me know if you hit any other missing nodes — happy to help! 🙌
Very detailed,well-formated, and thorough. Thanks.
Thanks Jeffka! Glad it's useful. Let me know if you run into any missing nodes or model path issues — happy to help 🙌
I wish you would share your samples with metadata so I would easily figure out the settings and try, still couldn't make it work but on it :l
also, camera is not locked in, slightly pushes in in your samples, so I don't know if static fully locked in camera in prompt work in here, or does it?
Good points! To clarify both:
On metadata:
The samples are MP4 videos — ComfyUI doesn't embed workflow metadata into video files, only into PNG images saved via the SaveImage node. Since this workflow outputs video directly through VHS, there's no metadata carrier.
All generation settings are documented in the description above, but here's a quick summary so you don't have to scroll:
- Sampler: er_sde | Steps: 8 base + 3 refine | CFG: 1.0
- Pixel budget: 0.52MP (auto-scaled per input) | Frames: 168 | FPS: 24
- Output: 960×544 → 2× spatial upscale → 1920×1088
- No LoRA, no ControlNet
On camera lock:
You're right — it's soft guidance, not a hard lock. LTX 2.3 doesn't have a rigid camera constraint system like ControlNet. The slight push-in happens because QwenVL detects scene depth and the model interprets it as subtle forward motion even when the prompt says "static."
To minimize it: after the workflow loads, find the QwenVL text output node and manually append this before queuing:
, static camera, no zoom, no pan, no push, locked perspective, fixed viewpoint
That tightens it noticeably. It won't be 100% rigid but the drift becomes much more subtle — closer to natural parallax than actual camera movement.
Hope that helps get it running!
@TP_AI_63 I think it is not because of Qwen, I tried without it too still didnt work. I cannot get static camera shot consistently. Maybe some if Im lucky.
Tried your workflow, your sampler settings are different then mine, now I'm trying those settings in my workflows to see the difference. Nice outputs btw, thank you.
@fakolonya You're right, the static camera issue is an LTX 2.3 limitation overall — the model just isn't great at fully rigid shots yet. The er_sde sampler + low CFG (1.0) does help reduce drift compared to default settings but it's never 100% consistent.
Good luck testing — hope the sampler settings make a difference for you! 🙌