DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Three Distillation Weights, Three Frame Geometries, One Release: Reading LTX 2.3's Configuration Drift

This release contains three workflows and, between them, three different distillation LoRA weights, three different frame counts, two different samplers, and a 2-bit quantised text encoder. That is not a criticism — it is what a real production toolkit looks like when each graph was tuned for its own job. But you have to read it carefully.

The three graphs

Workflow Nodes Geometry Distil. weight Sampler
Multi-image first/last frame 79 768×512, 97 frames 0.6 euler_ancestral_cfg_pp
IC LoRA Union video control 44 960×544, 121 frames 0.7 res_2s_ode
LTX 2 distilled lip-sync 106 768×512, 241 frames (LTX 2 LoRA) euler_ancestral

Three weights: 0.6, 0.7, and a different LoRA family entirely for the LTX 2 graph.

The 0.6 and 0.7 weights

Elsewhere in this family, the distillation LoRA appears at 0.5 (prompt-driven graphs) or 1.0 (conservative, reference-driven). Here it is at 0.6 and 0.7.

Those are intermediate values, and the reasoning is visible in the geometry:

  • 0.6 at 97 frames, 768×512 — moderate resolution, standard length. A bit more distillation than 0.5 because the base is larger and needs more efficiency.
  • 0.7 at 121 frames, 960×544 — larger base, longer segment. More distillation again, because the per-stage cost is higher and there is less budget to spend on undoing distillation.

The pattern: as the base geometry grows, distillation weight grows. That is a coherent rule, and it is more useful than picking a number.

The 2-bit text encoder

The multi-image graph uses DualCLIPLoaderGGUF with gemma-3-12b-it-Q2_K.gguf.

Q2_K is a 2-bit quantisation. For a 12-billion-parameter text encoder, that is aggressive. The payoff is memory: a 12B model at 2-bit is roughly a quarter the size of one at 8-bit.

The cost is instruction-following fidelity. A 2-bit Gemma 3 will follow obvious prompts and lose nuance — long prompts with multiple constraints degrade more than short ones.

The graph also loads the normal fp8 text encoder alongside, so the release is comparing or offering both. If prompt adherence matters, use the fp8 path. If memory is the constraint, Q2_K gets you running.

The 241-frame graph

LTX 2 distilled lip-sync uses:

EmptyLTXVLatentVideo → 768 × 512, 241 frames

241 frames at 24 fps is about ten seconds — two and a half times the 97-frame unit. And it needs LTXVEmptyLatentAudio at 121 frames.

Note the audio latent is roughly half the video frame count, which reflects the audio latent's own temporal resolution rather than a 1:1 mapping.

This graph uses LTX 2 (19B), not 2.3 — LTXVideo2\ltx-2-19b-distilled-fp8_transformer_only, with a ltx-2-19b-distilled-lora variant. So it is a different generation entirely, included because lip-sync at ten seconds is a job the older model still does well.

Where this release is weak

  1. Three weights with no documented rule. The pattern (larger geometry → more distillation) is inferable but not stated.
  2. Q2_K text encoder is a real quality tradeoff that is not flagged.
  3. Three geometries means no single tuning transfers between graphs.
  4. The 44-node IC LoRA Union graph uses res_2s_ode rather than the euler family, so sampling experience does not transfer.
  5. Two model generations in one release.

How to read a release like this

Do not look for a "recommended" graph. Look at the geometry first — 97 / 121 / 241 frames are three different working units for three different jobs. Then look at which model generation each graph uses. Then pick the one that matches your content's length and subject type.

Workflows

Three graphs, published on RunningHub with invite code rh-v1111:

  • LTX 2.3 multi-image first/last frame
  • LTX 2.3 IC LoRA Union video control
  • LTX 2 distilled lip-sync

Files

https://huggingface.co/AI-KSK/ksk-video-203-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts, geometries and weights quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with Lightricks or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)