LTX 2.3 Image to Video: Setup and Consistency Fixes

Image to video is the mode where AI video either earns your trust or loses it. You already have an image you like — a character, a product, a frame you spent time on. The only question that matters is whether the model can move it without wrecking it.

LTX 2.3 handles this through more than one route, and that is the part most guides flatten. There is a plain image-to-video path, a first-frame/last-frame path, a keyframe interpolation path with different conditioning math, and an in-context control path. They behave differently, need different checkpoints, and fail in different ways.

This guide covers all of them plus the constraints that silently break runs. Sourced from the official ComfyUI LTX-2.3 tutorial, the LTX-2 pipeline documentation, and the LTX-2.3 model card, checked July 2026.

Four ways to condition on an image

The technical distinction worth internalizing is in row three. The keyframe pipeline uses guiding latents — additive conditioning — instead of replacing latents. Replacement pins the model to your frame; addition nudges it. That is exactly why the interpolation route produces smoother transitions and the direct route produces stricter adherence to your input frame.

So: strict fidelity to the source image → I2V. Smooth motion between two states → keyframe interpolation or FLF2V.

Both two-stage pipelines also support image conditioning directly — TI2VidTwoStagesPipeline is the recommended production path and accepts an image alongside the text prompt.

The constraints that break runs before quality matters

Two hard rules from the model card, and they are not suggestions:

• Width and height must be divisible by 32.
• Frame count must follow the 8k+1 pattern — 97, 193, and so on.

Your source image almost certainly does not have dimensions divisible by 32. Crop or pad it before it enters the pipeline. A 1080x1920 portrait frame is fine (both divisible by 32); a 1000x1500 crop is not.

RetakePipeline restates the same constraints for its input video — frame count on 8k+1, resolution a multiple of 32 — which tells you these are model-level rules, not per-pipeline quirks.

One genuinely nice consequence of LTX 2.3's design: ComfyUI's documentation notes it handles native 9:16 portrait output without any cropping. For short-form work, that means the source image and the output share an aspect ratio and nothing gets letterboxed or center-cropped on the way through.

Why your subject drifts, and what to do about it

Face and character drift is the number-one complaint in image-to-video work. The causes are rankable, and most people attack them in the wrong order.

1. The source image is doing too little work. A small, soft, or partially-occluded face gives the model very little to preserve. Motion amplifies whatever ambiguity is already in the input. Fix the input first — it is free.

2. The prompt describes a scene instead of a continuation. With image conditioning, your prompt's job is not to re-describe what is already visible. It is to specify what happens next. A prompt that restates the image competes with the image; a prompt that describes motion complements it.

3. You are asking for too much time. Longer clips mean more frames for identity to wander across. Generate shorter, then extend — video_extend is a supported conditioning mode in the trainer's config set, and RetakePipeline lets you regenerate a bad window rather than a whole clip.

4. You need identity conditioning, not better luck. For a person who must be recognizably the same across many clips, the intended tool is ID-LoRA — ComfyUI ships a native ID-LoRA template using ltx-2.3-id-lora-talkvid-3k.safetensors, which does personalization with voice cloning. Prompt engineering does not substitute for identity conditioning.

5. You need structure, not identity. If what drifts is the pose or camera path rather than the face, that is IC-LoRA Union Control's job: structural guidance from depth, pose, or edges, driven by the moge_2_vitl_normal_fp16.safetensors geometry estimator.

Want to test how well your image survives? Before you build a local pipeline around a specific character, check how the model handles that image at all. ltx23.app runs LTX 2.3 image-to-video in the browser — a five-minute answer to a question that otherwise costs an evening of setup.

FLF2V: the mode people underuse

First-frame-to-last-frame gets treated as a niche feature. It is closer to a superpower for anyone doing commercial work, because it converts an open-ended generation problem into a constrained one.

You are not asking "what might happen"; you are asking "get from here to there." That is why it works so well for:

• Product shots that must start and end on an approved frame
• Logo and title transitions
• Morph-style effects between two designed states
• Any shot where a client already signed off on the endpoints

Note the checkpoint: ComfyUI's FLF2V template runs on the distilled FP8 build, not dev. That is a documented pairing, not a preference — loading the dev checkpoint into the FLF2V template is one of those silent mismatches that produces mediocre output with no error message.

A working image-to-video recipe

1. Prepare the source. Crop to dimensions divisible by 32. Prefer a clear, well-lit subject at reasonable scale in frame.
2. Pick the route. One image forward → I2V. Two fixed endpoints → FLF2V. Reference clip driving structure → IC-LoRA.
3. Match the checkpoint to the template (dev-fp8 for I2V; distilled-fp8 for FLF2V).
4. Write a continuation prompt. Describe motion, gesture, camera move, and what changes — not what the image already shows.
5. Set a legal frame count (8k+1) and keep the first pass short.
6. Iterate on distilled (8+4 steps), then render the keeper on the two-stage path with the spatial upscaler.
7. Repair, don't reroll. One bad second → RetakePipeline on that window, with independent regenerate_video / regenerate_audio control.

Step 7 is the one that separates people who ship from people who keep rerolling. A clip is not binary — a two-second problem in a ten-second clip is a two-second job.

FAQ

Which checkpoint does LTX 2.3 image to video use? In ComfyUI, the I2V template uses ltx-2.3-22b-dev-fp8; FLF2V uses ltx-2.3-22b-distilled-fp8.

What resolutions work? Anything with width and height divisible by 32. Native 9:16 portrait is supported without cropping.

Why does my video ignore the input image? Usually a route mismatch — additive keyframe conditioning nudges rather than pins. For strict adherence, use the direct image-conditioned path.

How do I keep the same character across clips? ID-LoRA. That is the mechanism built for it; prompting alone will drift.

Can I fix just one bad segment? Yes — RetakePipeline regenerates a specified [start_time, end_time] region and preserves the rest.

Can I extend a clip instead of regenerating longer? Video extension is a supported conditioning mode in the official trainer's config set, which is the intended path for longer material.

Bottom line

LTX 2.3 image to video is not one feature with one quality dial. It is four conditioning strategies, each with a checkpoint it expects and a failure mode it avoids. Choose by what you are holding — one image, two endpoints, or a reference clip — and respect the divisible-by-32 and 8k+1 rules before blaming the model.

To see how your own image behaves before committing to a local setup, run it through ltx23.app.

Sources

Verified July 2026 against primary documentation:

1. ComfyUI official LTX-2.3 tutorial — I2V, FLF2V, IC-LoRA and ID-LoRA templates and their checkpoints
2. LTX-2 available pipelines documentation — keyframe guiding latents, IC-LoRA constraints, retake behavior and constraints
3. Lightricks/LTX-2.3 — Hugging Face model card — resolution divisible by 32, 8k+1 frame rule, variant list
4. Lightricks/LTX-2 — official GitHub repository — pipeline overview and image conditioning support
5. LTX-2 pipeline selection guide — choosing between one-stage and two-stage
6. LTX-2 installation & CLI flags — running the image-conditioned pipelines
7. ltx-trainer configs — i2v and video-extension conditioning modes
8. LTX-2.3-22b-IC-LoRA-Union-Control — structural control for pose and depth
9. LTX-2 optimization documentation — step and memory settings affecting iteration
10. ComfyUI-LTXVideo — official integration repository
11. LTX-2 technical paper (arXiv 2601.03233) — conditioning architecture

Template-to-checkpoint pairings follow ComfyUI's official documentation at the time of writing and can change with new releases.