Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:<ul><li>☕ <a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Buy me a coffee on Ko-fi</a></li><li>💛 <a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">Sponsor on GitHub</a></li><li>🔁 <a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a> (recurring)</li><li>⚡ Or right here: the Civitai tip button on this page sends Buzz directly.</li></ul>A one-hop Q8_0 quantization of the TRUE base <code>google/gemma-3-12b-it</code>, in the loader format the JoyAI-Echo ComfyUI node pack expects. This is the exact text encoder the LTX-2.3 / JoyAI-Echo DiT family was trained against — nothing finetuned, nothing re-quantized.<h2 id="files">Files</h2><ul><li><code>gemma3-12b-BASE-clean-Q8_0.gguf</code> — 12.45 GiB, reference tier, ~0.55% RMS error vs bf16. Start here.</li><li><code>gemma3-12b-BASE-clean-Q4_0.gguf</code> — 7.31 GiB, half-size tier, embedding table held at Q8_0. Take this if those 5 GiB are what fit your card.</li></ul>Both are one-hop quantizations of the same clean bf16 base, in the same loader format.<h2 id="why-this-exists">Why this exists</h2>The commonly circulated <code>gemma3-12b-joyecho-Q8_0.gguf</code> is not base Gemma. Its architecture string is a different, modified model, and it was built from an fp8 checkpoint and then re-quantized — double quantization; it still carries <code>.comfy_quant</code> marker tensors from the fp8 stage. In my renders it produced duplicated subjects and identity corruption.The DiT reads base Gemma's embedding space. Any encoder drift degrades output structurally, not subtly — so this is one of the few places in the stack where the wrong file does not merely cost quality, it changes what you get.<h2 id="symptoms-of-the-wrong-encoder">Symptoms of the wrong encoder</h2><ul><li>Two of the same person in one frame, or a subject that splits Why: embedding drift from a modified or double-quantized encoder. Fix: use one of these files.</li><li>Prompt adherence stays vague no matter how you word it Why: the DiT is reading a slightly different embedding space than it was trained on. Fix: same — the encoder sits upstream of everything, so it is worth ruling out before you tune anything downstream.</li></ul><h2 id="everything-else-i've-published">Everything else I've published</h2><ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino">All models on Hugging Face</a></li><li><a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/joyai-echo-multishot-patch">JoyAI-Echo multishot patch</a> — nodes, releases, issues</li><li><a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/ltx23-multishot-lite">LTX-2.3 Multishot Lite</a> — the core-node workflow</li></ul>

Files with SHA256: 462cafdb13482793371520012d03f92dd30668c1a42d42890ec07fbaafd86a06