<h3 id="ref2va-workflow-ltxdirector-upscale">REF2VA Workflow LTXDirector Upscale</h3>I added a REF2VA workflow; it’s based on my I2V workflow but includes new changes and nodes. The issue with reference video is that, at higher denoise levels, LTX starts to eat into the original footage. To address this, I added LTXDirector to the upscaling pass, allowing me to incorporate references into the upscaling process. I’m no expert on LTXDirector—I don’t fully grasp all its features—but it gets the job done. I’ve also completely abandoned manual sigma settings; adjusting the denoise level makes it easier to control the final result. All the models remain the same as in the I2V workflow.<h3 id="important-changes:">Important changes:</h3><ol><li>I switched the generation preview to a node, just like in the Minimax pass, so copy the <code>taeltx2_3.safetensors</code> file to the path \ComfyUI\models\vae_approx.</li><li>I updated my node to display video data after loading—useful for editing in the LTX Director. If you're wondering why my node is needed at all, the answer is simple: it allows you to transfer metadata to another video, which is handy if you're upscaling the video in a different workflow, and so on. So, download and install it using the link <a rel="ugc" href="https://drive.google.com/file/d/14loLrzWWO7s95-EMNcjmul7j5Z3uHKe6/view?usp=drive_link">here</a> or below in the description.</li><li>Sigma is no longer available; use Denoise on the subgraph. For me, the ideal Denoise value is 0.4–0.45.</li><li>The same changes also apply I2V workflows in the new version 1.6.</li></ol><h3 id="minimax-h3-+-ltx-video-2.3-upscale-simple-low-vram-workflow-with-metadata">MiniMax H3 + LTX-Video 2.3 Upscale— Simple Low-VRAM Workflow with Metadata</h3>A clean and practical ComfyUI workflow designed for people with limited VRAM. It generates video using MiniMax H3 at a very low resolution (0.2 MP), then upscales the result with LTX-Video 2.3.You can run everything in a single pass or execute the stages one after another to keep GPU memory usage under control.<h3 id="features">✨ Features</h3><ul><li>🎬 Two-stage pipeline: MiniMax H3 (low-res) → LTX 2.3 (upscale)</li><li>🧠 Designed for low VRAM — first pass runs at only 0.2 megapixels</li><li>🖼️ First Frame + Last Frame support</li><li>⏱️ Easy video duration control in seconds</li><li>🪜 Adjustable sampling steps</li><li>🎨 Full LoRA support through Lora Manager</li><li>💾 Automatic metadata embedding into the final video</li><li>🔄 Option to run stages together or separately (bypass groups)</li></ul><h3 id="required-models">🤖 Required Models</h3>First Pass (MiniMax H3)<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3294059?fileId=3178732">Model </a>(or any compatible MiniMax H3 model)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors?download=true">Text Encoder</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors?download=true">Video VAE</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors?download=true">Audio VAE</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-TAE/resolve/main/vae_approx/taeh3.safetensors">Preview VAE </a>. \ComfyUI\models\vae_approx</li></ul>Upscale Pass (LTX 2.3)<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/CornLogic/10EROS_1.4_Int8_ConvRot/resolve/main/10Eros_v1.4_Bob_DMD_INT8_Convrot.safetensors?download=true">Model </a>(or any compatible LTX 2.3 model)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/resolve/main/gemma_3_12B_it_fp8_e4m3fn.safetensors?download=true">Text Encoder</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/text_encoders/ltx-2.3_text_projection_bf16.safetensors?download=true">Text Projection</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/vae/LTX23_video_vae_bf16.safetensors?download=true">Video VAE</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/vae/LTX23_audio_vae_bf16.safetensors?download=true">Audio VAE</a></li><li><a target="_blank" rel="ugc" href="https://github.com/madebyollin/taehv/raw/main/safetensors/taeltx2_3.safetensors">Preview VAE </a> \ComfyUI\models\vae_approx</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-spatial-upscaler-x2-1.1.safetensors?download=true">Spatial Upscaler</a></li></ul>The workflow is not locked to these specific models — you can swap them for other MiniMax H3 / LTX 2.3 variants.<h3 id="required-custom-nodes">🧩 Required Custom Nodes</h3>Most nodes can be installed through ComfyUI Manager.<h3 id="important!">Important!</h3>If you get an error related to Image Saver — Add Metadata to VHS Video, install ComfyUI-Image-Saver manually:Bash<pre><code>cd custom_nodes
git clone https://github.com/alexopus/ComfyUI-Image-Saver.git
cd ComfyUI-Image-Saver
pip install -r requirements.txt</code></pre>Then restart ComfyUI.<h2 id="you-need-also-install-my-custom-node.">You need also install my custom <a target="_blank" rel="ugc" href="https://drive.google.com/file/d/14loLrzWWO7s95-EMNcjmul7j5Z3uHKe6/view?usp=sharing">Node.</a></h2>Just extract folder vhs_metadata_loader into \ComfyUI\custom_nodes\ and restart ComfyUI.<h3 id="metadata-+-civitai">💾 Metadata + Civitai</h3>One of the main goals of this workflow is proper metadata preservation.The final video embeds:<ul><li>Prompt</li><li>Seed</li><li>Model information</li><li>LoRA information</li><li>Generation parameters</li><li>Other workflow metadata</li></ul><h3 id="civitai-video-metadata-assistant">Civitai Video Metadata Assistant</h3>To make use of the embedded metadata on Civitai, install the following browser extension:Last version you can find here, and don't forget to like the author's post:<a target="_blank" rel="ugc" href="https://civarchive.com/models/2898818/civitai-video-metadata-assistant-for-uploading-videos-browser-extension">https://civarchive.com/models/2898818/civitai-video-metadata-assistant-for-uploading-videos-browser-extension</a>Installation instructions are here:<a target="_blank" rel="ugc" href="https://github.com/diodiogod/Civitai-Video-Metadata-Assistant">https://github.com/diodiogod/Civitai-Video-Metadata-Assistant</a>After installation, the assistant will appear on the Civitai video upload page.It can read the metadata embedded in your generated video and automatically detect relevant information such as your LoRAs and generation data, making uploads much easier.<h3 id="how-to-use">🚀 How to use</h3>1 Pass<ol><li>Enable First Pass Group. Disable Last Frame if you don´t need them.</li><li>Write your prompt, Set video length (in seconds), set steps.</li><li>Choose resolution (0.2 MP recommended for first pass)</li><li>Add LoRAs if needed</li><li>Generate</li></ol>2 Pass<ol><li>Enable Upscale Pass Group. Disable Last Frame if you don´t need them. You need to enable/disable both togehter.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/48a18269-f1f0-44d7-afdf-affabb6f5c38/original=true/48a18269-f1f0-44d7-afdf-affabb6f5c38.jpeg" /></li><li>Load generated video with Metadata.</li><li>Copy your prompt from 1 Pass.</li><li>Add LoRAs if needed</li><li>Generate</li></ol>After bypassing groups, always check the status of the last frame, as it is turned on automatically.It is also possible to generate everything at once; to do this, you need to remove the node responsible for switching between the 1-pass and upscaling passes.<h3 id="why-this-workflow">❤️ Why this workflow?</h3>I am the "proud" owner of a system with 8GB of VRAM and 16GB of RAM. Minimax was forbidden fruit for me, but I found a workaround. Yes, the Ltx upscaler alters the video, but in my opinion, the result is still excellent.<h2 id="some-generation-data-on-my-hardware:">Some generation data on my hardware:</h2>10s video 0.1 MP. End video 512x7681 Pass 230s2 Pass 200s10s Video 0.2 MP. End Video 768x11521 Pass 470s2 Pass 458sPure LTX 2.3 10s Video with 640x1024 take 570sI spent a lot of time on this workflow, so I appreciate any feedback.<h3 id="notes">📌 Notes</h3><ul><li>Results depend on the models, LoRAs, source images, prompt and settings.</li><li>For best low-VRAM performance, keep the first pass at 0.2 MP.</li><li>You can disable groups to run the stages one by one if your GPU is struggling.</li></ul>

Simple Minimax H3 Fl2VA Workflow with LTX 2.3 Upscale and Metada save. + REF2VA Workflow LTXDirector Upscale(For low VRAM users) NSFW

The video shift has been adjusted.In the upscaling pass, the sampler has been changed to Euler, resulting in a noticeable speed boost. An option to choose between two sets of sigmas has been added: First allows for greater Ltx-Video influence— sharper quality and more significant detail changes.Second prioritizes fidelity to the original and offers a slight speed boost.

Files with SHA256: cd646c60ec68dc967820362064730b94c914d19c2b3b78fac7695add4065775d