ABERCROM-ME v3 — IC-LoRA for LTX-2.3-22B

Transform a photo of yourself (or a friend, or a couple) into a scene straight out of a vintage A&F Quarterly catalog — moody orthochromatic monochrome, sun-drenched beach or locker-room energy, and a full styling makeover to match. This is an image-conditioned LoRA (IC-LoRA): you feed it a reference photo and it restyles the subject, clothing, and background to match the aesthetic while preserving identity and pose.

This is version 3 — versions 1 and 2 never made it past my own testing, they just weren't good enough to put in front of anyone else. This one finally is.

Model File

ltx-2.3-22b-ic-lora-abercrom-me-3-2000.safetensors

Model Details


Base model	LTX-2.3-22B
Training type	IC-LoRA (image-conditioned, reference-latent)
Control signal	A single reference image of the subject
Trigger word	`ABERCROM-ME`
Recommended strength	1.0
Rank / Alpha	64 / 32
Target modules	Attention + FFN (`ff.net.0.proj`, `ff.net.2`)
Training steps	2,000
Resolution buckets	704×448×121 · 448×704×121 · 576×704×121 · 704×576×121 @ 24 fps

The reference image is looped into a static video and conditioned into the latent stream, so the model sees the subject throughout generation rather than only at the first frame. Identity and pose carry through; clothing, lighting, grade, and background are restyled.

Intended Use & Out-of-Scope

Intended use: Applying a vintage A&F-catalog editorial look to adults in short creative clips — an ordinary photo in, a styled orthochromatic monochrome scene out. Built for personal creative work, stylized portraiture, and fun.

Out-of-scope:

Photographic accuracy. This is a generative restyle, not a color grade. Clothing, background, and lighting are invented to fit the aesthetic. Nothing it produces is a faithful record of the source photo.
Minors. The model was trained exclusively on adult subjects with captions constrained to adult phrasing. Do not use it on images of children.
Real people without consent. Restyling a recognizable person produces plausible imagery of them in a place and outfit they were never in. Use it on yourself and on people who agreed to it.
General text-to-video. It expects reference-image conditioning. Without one, output is unpredictable.
Color output. Training targets were orthochromatic monochrome throughout. The model pulls hard toward that look and will fight attempts to keep a scene in full color.

Usage

Download the .safetensors file into models/loras.
Load the LTX-2.3-22B dev bf16 checkpoint and apply the LoRA at strength 1.0.
Supply a reference image of your subject.
Prompt with the trigger word:

ABERCROM-ME

Or add scene detail to steer the moment — the trigger sets the look, the rest of the prompt directs the action:

ABERCROM-ME two men in their twenties standing beside a wooden lifeguard
tower on a beach, orthochromatic silver gelatin editorial look, warm
vintage summer styling, confident relaxed motion.

Recommended Settings

Run against the dev bf16 checkpoint.

Setting	Value
Base checkpoint	LTX-2.3-22B dev bf16
ABERCROM-ME strength	1.0 (stage 1 only)
LTX distilled LoRA	0.25 in stage 1 · 0.6 in stage 2
Steps	20
CFG	3
Sampler	`euler_cfg_pp`
Schedule	`linear_quadratic`
Resolution	704-class buckets, or HD 1280×704 / 704×1280
Frames	8k+1 (121, 193, 481) @ 24 fps

Note that two LoRAs are in play: ABERCROM-ME itself, and the LTX distilled speed LoRA. ABERCROM-ME goes in stage 1 only at full strength. The distilled LoRA runs in both stages, but at different strengths — 0.25 for base generation, 0.6 for the spatial upscale.

Workflow

This follows the same IC-LoRA pattern as other reference-image LoRAs you may already use (Ingredients, MSR, etc.) — reference image in, styled result out.

Important: if you're running a 2-phase (base + upscale) pipeline, apply this LoRA only in the base generation stage. Do not apply it during the stage-2 upscale pass — that's the one thing that trips people up coming from a standard workflow.

My ComfyUI workflow is included in the repo files. The only things you need to set are a starting image and your model paths. Check the note nodes on the far left for the custom nodes it depends on (Pixaroma nodes and my node pack).

Three stages:

Stage 1 — base IC-LoRA generation
Stage 2 — LTX spatial upscale
Stage 3 — LTX temporal upscale, doubling framerate

There is no other published ComfyUI workflow for the temporal upscale model, ltx-2.3-temporal-upscaler-x2-1.0.safetensors, so I'd genuinely like to hear how stage 3 behaves for you. Skip it if it gives you trouble — it looks great at 24 fps regardless.

The workflow uses Ollama to draft or enhance a prompt, and always prepends the ABERCROM-ME trigger automatically.

Showcase

Beach (1024×1280, 193 frames)

ABERCROM-ME The two men stand on the beach as gentle waves roll in behind them. Suddenly, a bright flash of sunlight breaks through the clouds above, illuminating their faces and casting long shadows across the sand. A seagull swoops down from the sky, gliding low over the water before circling back up into the light. The man in the blue shirt smiles faintly as he watches it pass.

Cliff (704×1280, 121 frames)

ABERCROM-ME The man, shirtless and wearing dark swim shorts, takes a confident step forward while turning his head slightly toward the camera. He raises one arm flexing as he says in an energetic tone, "Let's go!" A sudden burst of light reflects off the mirror behind him as if signaling something exciting has begun.

Storm (1280×704, 193 frames)

ABERCROM-ME The man stands still, gazing into the distance as a sudden gust of wind sweeps across the field, rustling the tall grass and sending a flurry of sand spiraling through the air. A low rumble builds in the sky, growing steadily louder—then with a sharp crack, lightning splits the clouds overhead, followed by a deep roll of thunder that echoes across the open landscape.

Lifeguards (1024×1280, 193 frames)

ABERCROM-ME two men in their twenties standing beside a wooden lifeguard tower on a beach, orthochromatic silver gelatin editorial look, warm vintage summer styling, confident relaxed motion.

Lake (1280×704, 481 frames)

ABERCROM-ME orthochromatic style, The two men stand beside their pickup truck on the gravel lakeshore, the lake calm beyond and tree-lined hills rising in soft focus. The camera slowly pushes forward between them toward a small fishing boat drifting in the distance. Suddenly, the lake swells beneath the boat and a massive sea monster surfaces, snapping the boat in half. The man on the left exclaims "Oh crikey!" and the other replies, "Well this fantasy was fun, but Abercrom ME outta here!" They turn and sprint away together across the gravel.

Mechanic (832×1280, 193 frames)

ABERCROM-ME A man stands beside a red truck with its hood open on a dirt road, leaning forward as he examines the engine. Suddenly, the engine sputters to life, emitting a deep, resonant rumble. The man's face lights up with relief and excitement as he steps back, nodding in satisfaction.

Tips and Limitations

Clean reference images work best. A clear subject against an uncluttered background gives the strongest, most consistent restyle.
Busy scenes drift. Many figures in frame means uneven style application; some subjects get the full treatment while others lag.
POV and extreme close-ups underperform. With little visible clothing or environment to work with, there's not much for the model to restyle.
Grayscale references help. A low-saturation or grayscale reference image reaches the target look faster. A saturated color first-frame anchor fights the orthochromatic grade.
Seed matters. Results vary meaningfully run to run; it's worth generating a few.
Video appearance only. This affects the visual style. Audio is untouched.

Dataset

52 paired samples, built specifically for reference-conditioned style transfer.

Target videos were generated with I2V workflows across LTX-2.3 and Wan 2.2, starting from images produced with an Abercrombie-style LoRA trained for Krea 2. Targets were meant to be orthochromatic throughout; any clips that drifted were corrected with a Python script built on an orthochromatic-conversion library to bring them back in line.

To create the conditioning input for each pair, the first frame of the target video was run through GPT Image to generate an "everyday" version of the same person — stripping the styling back to an ordinary photo. Each pair is therefore a plain reference photo in, and a full A&F-styled orthochromatic monochrome video out, with the model learning that exact transformation.

Scenes span beaches, lakeshores, pickup trucks, locker rooms and open fields, in both landscape and portrait orientation. Captions follow one consistent shape: trigger word, then scene, then motion, one sentence per clip. Age language was constrained deliberately across the whole set — always "a man in his twenties" or "adult men," never ambiguous phrasing — and every caption was validated against that rule before training.

Training

IC-LoRA with reference-image latent conditioning on LTX-2.3-22B, trained with the LTX Video Trainer.

Rank 64, alpha 32, targeting attention plus FFN (ff.net.0.proj, ff.net.2). Trained from base for 2,000 steps, int8-quantized during training, across four 704-class resolution buckets at 121 frames.

License

Model weights are released under the LTX-2 Community License.

Feedback Welcome

This is a first public release, and I'd genuinely appreciate feedback on where it holds up and where it breaks — odd clothing artifacts, identity drift, background coherence, that kind of thing. It shapes where v4 goes.

Acknowledgments

Base model and training infrastructure by Lightricks, via the LTX-2 Community Trainer. IC-LoRA adapter trained by CoachBate.

Downloads last month: 255