Light Being Forest Reveal — ComfyUI + LTX Video 2.3 — local workflow notes
A test in pushing LTX 2.3 toward a hand-drawn woodcut aesthetic, run entirely on local hardware. The figure removes its skin and reveals a light being inside. The cat witnesses.
Local models are a real game-changer for me. Hand-drawn is my style — I want to animate my own artworks without depending on expensive cloud setups or paying per generation to experiment. Every iteration on a hosted service is a credit burned. On local, the only cost is time. The experiments get longer, weirder, and more honest. This is the first one I want to share end-to-end: prompt, workflow, settings, hardware, and what I learned.
ComfyUI + LTX 2.3 Starting Image
PROMPT
A flat comic-book illustration of a figure in a pink jumpsuit with a smooth featureless white helmet, standing motionless in a forest clearing at night. Above them, a mystical luminous eye-mandala hovers in the starry sky with radiating beams of light. Two small black cats sit symmetrically at the figure's feet. Tall teal-green pine trees flank the scene. High detailed hand-drawn comics style — clean ink linework, stippled shading, flat saturated colors in pink, teal, yellow, and deep black.
The figure begins to unzip and remove the pink jumpsuit. The helmet lifts and tilts away first, revealing that the head underneath is a source of bright warm white-yellow light — a radiant glowing orb where the head should be, bleeding soft lens glow into the surrounding air. The jumpsuit then peels open down the front, and from the opening more light pours out — the entire body beneath the suit is a luminous figure made of pure warm light, humanoid in shape but composed of glowing energy, with no skin or flesh, only radiance. The pink fabric of the suit falls away slowly, drifting to the ground in loose folds, as the being of light steps forward out of it, body softly pulsing and emanating rays that echo the mandala above. The two black cats look up, alert and watching. The mystical eye-mandala in the sky responds — its beams pulse brighter, resonating with the figure below, as if recognizing kin. The stars twinkle. The pine trees remain still. Static locked-off camera, no camera movement, no zoom, no pan. The hand-drawn comics illustration style is fully preserved throughout the entire transformation — visible ink lines, flat color fills, stippled texture on every surface, the glowing light rendered as radiant flat yellow-white gradients with linear beam rays in comic book style, not photorealistic glow. Cross-hatching linework, dark-anime atmospheric mood, ritualistic revelation, mystical folk-cosmic illustration, stippled shading, limited palette, (meditative transcendence.:1.1)
ComfyUI + LTX 2.3 Local Workflow
Model: LTX Video 2.3 (22b distilled, 1.1) Upscaler: ltx-2.3-spatial-upscaler-x2 (1.1) Text encoder: Gemma 3 12B IT (fp4 mixed) + LTX 2.3 text projection Video VAE: LTX23_video_vae_bf16 Audio VAE: LTX23_audio_vae_bf16 (CPU) Preview VAE: taeltx2_3 (Tiny VAE) Resolution: 1280 × 832 Length: 8 seconds FPS: 30 Total frames: 240 Sampler: euler_ancestral_cfg_pp Scheduler: LTXVScheduler with ManualSigmas CFG: 1.0 NAG scale: 7.5 NAG alpha: 0.25 NAG tau: 2.5 Image strength: 0.9 (LTXVImgToVideoInplace) Img compression: 22 Seed: 43 (fixed) Hardware: MacBook Pro M1, 64GB unified memory Generation time: ~84 minutes
THE TECHNIQUE I LEARNED— CFG 1.0 + NAG
LTX 2.3 use something called CFG values. When I push at 4, I get over-correction and the stylized aesthetic collapses into something realistic. I lowe the numbers, the prompt loosens and the model freelances. The sweet-spot I found for this style: drop CFG to 1.0 and let NAG (Negative Augmentation Guidance) carry the negative conditioning at much higher scale.
My settings: CFG: 1.0 nag_scale: 7.5 nag_alpha: 0.25 nag_tau: 2.5
What this does in practice: the positive prompt is interpreted loosely (which preserves stylistic interpretation), but the negative prompt is enforced hard (which blocks the realism default). The result is a generation that holds its woodcut character all the way through the transformation. Without NAG, I get much looser, more generic results — the character become a 3d light, CGI effect within a few frames. With NAG at these settings, the linework survives the transformation.
LTX 2.3 will smooth out woodcut linework if you don't fight it. The prompt lthat worked better leans hard on negation → "NOT digital, NOT vector, NOT smooth" to keep the halftone grain alive through generation.
The workflow above is one experiment. The reason I run experiments like it: I'm building a serialized cosmic horror universe called Atlas K, released as short films on YouTube.
Each "Fragment" is a 5–7 minute film built end-to-end with local AI tools and hand-drawn source aesthetics.
If you want to see what these workflows produces at full length: → Atlas K on YouTube: youtube.com/@Kajimelo
The channel is growing steadily. Subscribing helps it grow faster.
Written from Bolzano, Italy, where I run a one-person AI animation studio. Available for selective commercial work — AI animation, short film direction, AI-assisted advertising. Get in touch: kajimelo.com/contact
