{"id":5954,"date":"2026-03-27T17:26:55","date_gmt":"2026-03-27T09:26:55","guid":{"rendered":"https:\/\/crepal.ai\/blog\/?p=5954"},"modified":"2026-03-27T17:26:57","modified_gmt":"2026-03-27T09:26:57","slug":"ltx-2-3-multi-stage-latent-upscaling-comfyui","status":"publish","type":"post","link":"https:\/\/crepal.ai\/blog\/aivideo\/ltx-2-3-multi-stage-latent-upscaling-comfyui\/","title":{"rendered":"LTX 2.3 Multi-Stage Latent Upscaling Workflow in ComfyUI"},"content":{"rendered":"\n

Hi there, this is Dora. Two weeks ago I watched a 5-second clip I’d generated at 480p look genuinely cinematic after running it through the LTX 2.3 multi-stage latent upscaling pipeline<\/a>. Same prompt. Same motion. Completely different level of sharpness and edge detail. I actually said “wait, what?” out loud to my empty room.<\/p>\n\n\n\n

LTX 2.3 dropped on recently and the ComfyUI-LTXVideo repository<\/a> shipped with reference workflows for multi-stage latent upscaling on day one \u2014 but the documentation assumes you already know what latent upscaling is and why it matters. Most creators don’t. This guide fixes that.<\/p>\n\n\n\n

What Multi-Stage Latent Upscaling Is (Concept in Plain Language)<\/h2>\n\n\n\n

Standard video generation works like this: you give the model a prompt, it generates a video at whatever resolution you asked for, done. Single pass. One resolution. What you see is what you get.<\/p>\n\n\n\n

Multi-stage latent upscaling is a different approach. Instead of generating at full resolution in one shot, you:<\/p>\n\n\n\n

    \n
  1. Generate at a lower resolution in latent space<\/strong> \u2014 getting motion structure, scene coherence, and temporal consistency right first<\/li>\n\n\n\n
  2. Upscale within the latent space<\/strong> before decoding \u2014 adding spatial detail without regenerating the whole clip<\/li>\n\n\n\n
  3. Run a second denoising pass<\/strong> on the upscaled latent to lock in fine texture, edge sharpness, and lighting detail<\/li>\n\n\n\n
  4. Optionally<\/em>, apply a final pixel-space upscale for maximum output resolution<\/li>\n<\/ol>\n\n\n\n
    \n
    \"\"<\/figure>\n<\/figure>\n\n\n\n

    The key word is latent<\/em>. You’re not upscaling the decoded video frames (that’s pixel-space upscaling, like Topaz Video AI does). You’re operating directly on the compressed latent representation \u2014 the model’s internal “understanding” of the video \u2014 before it ever gets turned into pixels. This preserves temporal coherence across frames in a way that pixel-space upscaling cannot.<\/p>\n\n\n\n

    The result: sharper detail, preserved motion consistency, and significantly better edge accuracy \u2014 especially on fine textures like hair, fabric, and text \u2014 compared to generating at the target resolution in a single pass.<\/p>\n\n\n\n

    Why Use Multi-Stage vs. Single-Pass Generation<\/h2>\n\n\n\n

    The honest answer: single-pass generation at high resolution is computationally expensive and temporally inconsistent. Here’s why the two-stage approach wins:<\/p>\n\n\n\n

    Factor<\/td>Single-Pass (High Res)<\/td>Multi-Stage Latent Upscale<\/td><\/tr>
    Motion coherence<\/td>Can drift on complex motion<\/td>\u2705 Established at lower res Stage 1<\/td><\/tr>
    Fine detail<\/td>Present but can over-constrain motion<\/td>\u2705 Added in Stage 2 upscale pass<\/td><\/tr>
    VRAM at generation<\/td>High (full resolution throughout)<\/td>Lower (base res in Stage 1)<\/td><\/tr>
    Temporal consistency<\/td>Risk of frame-to-frame drift<\/td>\u2705 Preserved across upscale<\/td><\/tr>
    Generation speed<\/td>Slower for equivalent detail<\/td>Faster Stage 1 + targeted Stage 2<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    The community consensus on X and Reddit, backed by Lightricks’ own blog documentation<\/a>, confirms that the two-stage pipeline consistently outperforms single-pass at equivalent compute. The gap is most visible on shots with complex texture \u2014 skin, fabric, foliage \u2014 and on clips longer than 4 seconds where single-pass generation starts to drift.<\/p>\n\n\n\n

    One important constraint worth knowing upfront: width and height settings must be divisible by 32 at every stage. This trips people up when setting custom resolutions.<\/p>\n\n\n\n

    The Official Multi-Stage Workflow Walkthrough<\/h2>\n\n\n\n

    Before touching any node settings, your file structure needs to be correct. Here’s the required directory layout for the two-stage pipeline:<\/p>\n\n\n\n

    ComfyUI\/\n\u251c\u2500\u2500 models\/\n\u2502   \u251c\u2500\u2500 checkpoints\/\n\u2502   \u2502   \u2514\u2500\u2500 ltx-2.3-22b-dev-fp8.safetensors      # or BF16 if VRAM allows\n\u2502   \u251c\u2500\u2500 latent_upscale_models\/\n\u2502   \u2502   \u2514\u2500\u2500 ltx-2.3-spatial-upscaler-x2-1.0.safetensors   # required for Stage 2\n\u2502   \u251c\u2500\u2500 loras\/\n\u2502   \u2502   \u2514\u2500\u2500 ltx-2.3-22b-distilled-lora-384.safetensors    # required for pipeline\n\u2502   \u2514\u2500\u2500 text_encoders\/\n\u2502       \u2514\u2500\u2500 gemma_3_12B_it_fp4_mixed.safetensors<\/code><\/pre>\n\n\n\n
    The Spatial Upscaler and Distilled LoRA are both required for current two-stage pipeline implementations in the ComfyUI-LTXVideo repository. If either file is missing, the workflow will fail silently or produce artifacts at the upscale stage.<\/p>\n\n\n\n
    \n
    \"\"<\/figure>\n<\/figure>\n\n\n\n
    Stage 1: Base Generation Settings<\/h3>\n\n\n\n
    Stage 1 focuses entirely on motion structure and scene coherence<\/strong> \u2014 not on detail. You’re generating at roughly half your target resolution.<\/p>\n\n\n\n
    Target resolution for Stage 1:<\/strong><\/p>\n\n\n\n
    Final Output<\/td>Stage 1 Resolution<\/td>Notes<\/td><\/tr>
    1280\u00d7720 (720p)<\/td>640\u00d7384<\/td>Divisible by 32 \u2705<\/td><\/tr>
    1920\u00d71080 (1080p)<\/td>960\u00d7544<\/td>Divisible by 32 \u2705<\/td><\/tr>
    2560\u00d71440 (1440p)<\/td>1280\u00d7720<\/td>Divisible by 32 \u2705<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    Key nodes in Stage 1:<\/strong><\/p>\n\n\n\n