JoyAI-Echo's performance on LTX-2.5's engine. LTX-2.5 makes picture and sound in one pass, at any length, in one generation. JoyAI-Echo - a fine-tune of LTX-2.3 - is the better actor: natural lip-sync, expressive faces, a voice that stays put. The two transformers are shape-identical, so JoyAI-Echo's video attention / feed-forward delta was transplanted onto the official LTX-2.5 dev transformer and the official LTX-2.5 distilled LoRA baked in at 0.5. Few-step files: same speed, same VRAM, same nodes as LTX-2.5 distilled. Nothing retrained.v2. The first build put the delta straight onto the distilled transformer and came out over-saturated, with hard contrast and blown highlights - measured +16% saturation and 2-5x the clipping of stock on the same seed, obvious on any sunlit scene. Those files are withdrawn; every file here is the dev-base rebuild. The plain dev merges - for applying your own distilled LoRA at your own strength - are a separate page, Joy-LTX 2.5 DEV Merges, and on Hugging Face (<a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-dev">joyai-echo-ltx25-echoVid-dev</a>).<h2 id="what-you-get">What you get</h2>LTX-2.5 distilled is fast and clean but a flat performer; this merge keeps its speed and adds the thing Echo was trained for - mouths that shape the words, faces that move with the line, a voice that stays the same person across a long take. It runs in the stock LTX-2.5 graph, and in the <a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/ComfyUI-JoyLTX25">ComfyUI-JoyLTX25</a> canvases (Civitai: Joy-LTX 2.5 workflows): a one-prompt Take sized to your card by a planner, and a Multishot that joins shots seamlessly at the VRAM of one shot. A 2-minute four-shot take rendered with it was described blind as "a single continuous take" with six seconds of dead air.<h2 id="two-doses">Two doses</h2><ul><li>070T30 (default) - 0.7 x Echo on video attention / FF, 0.3 x on the modulation tables, distilled LoRA 0.5. Natural grade, cleaner skin. Start here. (The files on this page.)</li><li>100T50 (strong) - 1.0 x / 0.5 x, distilled LoRA 0.5. Livelier on loud, comic, animated performances; a touch hotter on contrast. Same file names with <code>100T50</code>, on Hugging Face.</li></ul><h2 id="pick-a-version-by-your-card-then-a-file-by-your-gpu-family">Pick a version by your card, then a file by your GPU family</h2>Every version holds two files for the same tier - one GGUF, one comfy-native. Same model; the format is the only difference.<ul><li>RTX 30 / 40 (Ampere, Ada): take the <code>.gguf</code>. Measured on a 3090, Q5_K_M / Q6_K are 4-8x faster than any 4-bit comfy-native file. Needs ComfyUI-GGUF installed (the JoyLTX canvas loader takes .gguf and .safetensors alike; in a stock graph use <code>Unet Loader (GGUF)</code>).</li><li>RTX 50 (Blackwell): take the <code>comfy-*.safetensors</code>. Stock ComfyUI 0.32+, plain <code>Load Diffusion Model</code>. int8 / w4a8 / w4a4 / nvfp4 / mixed 4-8 all land at roughly 90-120 s per 8 s clip on a 5090; add <code>--enable-triton-backend</code> to the launch line or they run the eager fallback at about half speed.</li></ul><ul><li>16 GB cards (and 12 GB) - <code>Q4_K_S.gguf</code> (12.9 GB) / <code>comfy-w4a8</code> (12.5 GB); 12 GB cards: <code>Q3_K_M.gguf</code> (10.6 GB); <code>comfy-nvfp4</code> for RTX 50.</li><li>24 GB cards - <code>Q5_K_M.gguf</code> (15.9 GB, the 24 GB default) / <code>Q6_K.gguf</code> (17.7 GB) / <code>comfy-mix4x8-17.0GB</code> (RTX 50).</li><li>32 GB cards - <code>comfy-int8</code> (21.5 GB, fastest on RTX 50) / <code>Q8_0.gguf</code> (22.7 GB).</li></ul>Everything, both doses, all quants (Q2_K to Q8_0; int8, w4a8, w4a4, nvfp4, mixed 13.8 and 17.0 GB): <a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-gguf">GGUF repo</a> · <a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-comfy-native">comfy-native repo</a>.<h2 id="you-also-need-(lightricksltx-2.5-on-hugging-face)">You also need (Lightricks/LTX-2.5 on Hugging Face)</h2><ul><li><code>ltx-2.5-video-vae-bf16.safetensors</code> and <code>ltx-2.5-audio-vae-bf16.safetensors</code> → <code>models/vae/</code></li><li><code>ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors</code> → <code>models/latent_upscale_models/</code> (optional for 24 GB: <code>ltx-2.3-spatial-upscaler-x1.5-1.0.safetensors</code> from LTX-2.3)</li><li>text encoder <code>gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors</code> → <code>models/text_encoders/</code> (16 GB cards: the 10.6 GB <code>gemma4-12b-ltx25-comfy-w4a8.safetensors</code> from my LTX-2.5-Quantized page)</li><li>the workflows: <a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/ComfyUI-JoyLTX25">ComfyUI-JoyLTX25</a> (Manager → Install via Git URL, or the zip on the Joy-LTX 2.5 workflows page - the writer with our prompts is inside it). No writer? paste your own prompt.</li></ul><h2 id="dials-worth-knowing">Dials worth knowing</h2><ul><li>video_cfg - 1.0 is the distilled default. The v2 files still sit a notch warmer than stock LTX-2.5 (saturation tracks stock, contrast +7-30% on the same seed); 0.7-0.85 brings them level, at the cost of the negative pass (~1.7x time). Same dial in the stock graph's guider.</li><li>Long takes - 30 s single pass at 1280x736 fits 24 GB; the Multishot canvas's identity anchor at 0.6 holds a 4 x 30 s take's texture flat (measured 1.00 0.88 1.11 1.10 vs 1.00 1.23 1.42 1.59 without).</li><li>Say the sounds - LTX invents drones and music for anything left unsaid; name the room tone. Say "American voice" if you want one.</li></ul><h2 id="numbers">Numbers</h2><ul><li>8 s at 960x544, two passes to 1920x1088: ~2 min on a 5090 (int8), ~7 min on a 3090 (Q5_K_M).</li><li>30 s single pass at 1280x736: 5 min (5090) / 17 min (3090). 12 s two-pass x1.5 on a 3090: ~7 min.</li><li>Multishot: ~2 min per 8 s shot on a 5090, ~4 min on a 3090; a 4 x 30 s take ~40 min on a 3090.</li></ul><h2 id="what-is-verified-and-what-is-not">What is verified, and what is not</h2><ul><li>Verified: every quant on both repos rendered (5 s and 10 s, high-key exterior and interior prompts, same seed against stock LTX-2.5 distilled); 2-minute takes on a 5090 (int8) and 3090 (Q5_K_M); identity across cut scenes; joins below motion.</li><li>Known: the warmer grade above; Q2_K is soft and hazy (it is a 2-bit); Q3_K_M is the floor for faces.</li><li>Not claimed: lip-sync scores - reviewers scored the merge and stock within noise of each other on an 8 s talking head; the difference is in the acting you can see, not in a number.</li></ul><h2 id="credits">Credits</h2>JoyAI-Echo by JD (jdopensource/JoyAI-Echo). LTX-2.5 by Lightricks. Merge, quantisation, nodes and canvases by me. Licensed under the LTX-2.x Community License (inherited from both parents). Questions: comment here or open a discussion on the Hugging Face page - I answer.

Joy-LTX 2.5 Distilled GGUF / INT8 / NVFP4 / W4A8 / MIX4x8

For 16 GB cards (and 12 GB with the Q3_K_M). Pick by GPU family:RTX 30 / 40 - GGUF (needs ComfyUI-GGUF, <code>Unet Loader (GGUF)</code>)<ul><li><code>LTX25dist-echoVid-070T30-v2-DiT-Q4_K_S.gguf</code> - 12.9 GB. The 16 GB default. 8 s clip ~8 min on a 3090.</li><li><code>LTX25dist-echoVid-070T30-v2-DiT-Q3_K_M.gguf</code> - 10.6 GB. For 12 GB cards; still looks right.</li></ul>RTX 50 - stock ComfyUI 0.32+, plain Load Diffusion Model<ul><li><code>LTX25dist-echoVid-070T30-v2-DiT-comfy-w4a8.safetensors</code> - 12.5 GB. The 16 GB default on Blackwell (~90 s per 8 s clip on a 5090). Also runs on 30/40 series but slowly - use the GGUF there.</li><li><code>LTX25dist-echoVid-070T30-v2-DiT-comfy-nvfp4.safetensors</code> - 12.5 GB. Blackwell only.</li></ul>Text encoder for 16 GB: <code>gemma4-12b-ltx25-comfy-w4a8.safetensors</code> (10.6 GB, on my LTX-2.5-Quantized page). Workflows: ComfyUI-JoyLTX25 (Manager). The 100T50 dose twins are on Hugging Face.

Files with SHA256: 11c515f0d9efb69e3805004369314be98e93e611376fac17c8d0df95241997d2