DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

A 19B LoRA on a 22B Checkpoint: What That Mistake Costs, and When It Does Not Matter

The smallest graph in this release loads a LoRA built for a 19B model on top of a 22B checkpoint. That is the kind of detail you would normally skim past — and it is exactly the kind of detail that explains otherwise inexplicable output quality.

The two graphs

Workflow Nodes Notable
Looped sampling upscale optimisation 45 ltx-2-19b-distilled-lora-384 at 1.0
Video control HD enhancement 80 960×544, 121 frames

The mismatch

The upscale graph loads:

  • CheckpointLoaderSimple → ltx-2.3-22b-distilled.safetensors (22B)
  • LoraLoaderModelOnly → ltx-2-19b-distilled-lora-384.safetensors (19B) at weight 1.0

Two different model generations. The checkpoint is LTX 2.3 at 22B; the LoRA is from the earlier LTX-2 at 19B.

These are not the same architecture. A LoRA is a set of low-rank deltas applied to specific weight matrices; if the target matrices do not match in shape and role, the deltas are applied to the wrong weights.

What actually happens in practice: the LoRA loads without error if tensor shapes happen to be compatible (a rank-r delta on a matrix can often apply even when the surrounding model differs), and the effect is somewhere between "partial" and "wrong". You get output, and you cannot tell whether the effect you see is the LoRA's intended contribution or noise.

If your upscale result looks subtly worse than expected, check the LoRA version first.

Why this matters beyond one workflow

The LTX family has, at this point in the timeline:

  • LTX-2 (19B) with ltx-2-19b-* LoRAs
  • LTX 2.3 (22B) with ltx-2.3-22b-* LoRAs
  • Multiple checkpoint variants within 2.3 (distilled, distilled-1.1, dev-dare-ties-distilled, dev-dare-merged-distilled, plus transformer_only fp8 variants)

Cross-loading between any of these is a real risk, and nothing validates it. The naming is close enough to skim past and different enough to matter.

The euler sampler

The upscale graph uses plain euler, not euler_ancestral_cfg_pp.

For an upscale pass that is defensible: upscaling should refine, not re-roll. Ancestral samplers inject noise at each step; for a refinement pass you usually want a deterministic trajectory that stays close to the input.

If your upscale produces something that looks like a different image rather than a sharper one, ancestral noise is a likely cause — and this graph correctly avoids it.

The 121-frame, 960×544 geometry

The HD enhancement graph uses a larger base than most LTX 2.3 workflows:

EmptyLTXVLatentVideo → 960 × 544, 121 frames

Versus the more common 320×288 / 97 frames. Higher starting resolution means the upscale path has less work to do and the whole graph can be shorter — which matches the 80-node count for a graph doing video control plus enhancement.

Where this release is weak

  1. The 19B/22B LoRA mismatch — the headline problem.
  2. ltx-2-19b-distilled-lora-384 on ltx-2.3-22b-distilled suggests the graph was built before the 2.3 LoRAs existed, and was not updated.
  3. 45 nodes for an upscale graph is thin. No internal validation.
  4. 121-frame geometry is non-standard for this family, so sigma schedules and sigma-derived settings from other releases do not transfer.

Practical use

Fix the LoRA first — swap to ltx-2.3-22b-distilled-lora-384 (the matching 22B version) and compare. That comparison is the most valuable thing you can do with this workflow: it tells you how much of your output quality was the mismatch.

Then treat the HD enhancement graph as a reference for the 960×544 / 121-frame configuration, which is a useful alternative geometry when 97 frames is too short.

Workflows

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

  • LTX 2.3 looped sampling upscale optimisation
  • LTX 2.3 video control HD enhancement

Files

https://huggingface.co/AI-KSK/ksk-video-216-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts and loader values 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 (1)

Collapse
 
alexshev profile image
Alex Shev •

Checkpoint and adapter compatibility deserves a preflight check rather than discovery through degraded output. I’d validate architecture family, layer shapes, tokenizer or text-encoder assumptions, and key coverage before loading, then surface a compact compatibility report. It makes an intentional partial merge distinguishable from a silent mismatch that only appears as strange generations.