dispatch-style-lora-ltx2-3/info.txt

═══════════════════════════════════════════════════════════════════════════════
  DISPATCH STYLE LORA LTX2.3
═══════════════════════════════════════════════════════════════════════════════

───────────────────────────────────────────────────────────────────────────────
  IDENTITY
───────────────────────────────────────────────────────────────────────────────
  Base model        LTX-2.3
  Generation mode   not declared by the author
  Checkpoint        not determined — verify before use
  Category          Visual style
  Author            tazmannner379
  Source            https://civitai.com/models/2375591
  Provenance        CivitAI · model 2375591 · version 2776562

───────────────────────────────────────────────────────────────────────────────
  WEIGHTS
───────────────────────────────────────────────────────────────────────────────
  File              dispatch23_v1_2-step00031000.comfy.safetensors
  Version           LTX 2.3 V1.1
  Size              1,285.6 MB
  Format            SafeTensor / unknown precision
  SHA256            54c9119f9f9f41dca49a00ff734c0020a90b38d1e15e62bca77ed50b84d2e4b6
  Published         2026-03-16

───────────────────────────────────────────────────────────────────────────────
  TRIGGER WORDS
───────────────────────────────────────────────────────────────────────────────
  › "DISPSTYLE"
  › "char_rr"
  › "char_bb"
  › "char_case"
  › "char_invisi"
  › "char_prism"
  › "char_malev"
  › "char_roy"
  › "char_punchup"

───────────────────────────────────────────────────────────────────────────────
  RECOMMENDED SETTINGS
───────────────────────────────────────────────────────────────────────────────
  The author did not publish specific settings.
  No author settings were published. Community baseline: steps Distilled path
  (all six official ComfyUI LTX-2.3 templates): 8 steps in stage 1 driven by
  an explicit ManualSigmas list "1.0, 0.99375, 0.9875, 0.98125, 0.975,
  0.909375, 0.725, 0.421875, 0.0", then 3 steps in the stage-2 refine after
  the x2 latent upscale ("0.85, 0.7250, 0.4219, 0.0") — verified by reading
  the template JSONs. The ic_lora template instead uses a plain KSampler at 8
  steps. The Python DistilledPipeline is documented as "8 predefined sigmas
  (8 steps in stage 1, 4 steps in stage 2)". Full/dev model with real
  guidance: Lightricks' own single-stage example workflow sets LTXVScheduler
  to 15 steps., cfg 1.0 for anything running the distilled checkpoint or the
  distilled LoRA — CFGGuider is set to 1 in both stages of every official
  ComfyUI template (verified in the JSONs), the ic_lora template's KSampler
  is cfg 1, and a note in Lightricks' own two- stage workflow reads "Do
  explore various samplers and cfg values (although we advise them to be kept
  close to 1)". For the guided full-model path, Lightricks' single-stage
  workflow uses MultimodalGuider with per-modality GuiderParameters: VIDEO
  cfg 3 (rescale 0.9), AUDIO cfg 7 (rescale 0.7). Treat 3 as the video anchor
  since that is the official value., sampler euler with SamplerCustomAdvanced
  + ManualSigmas is the default in the ComfyUI templates; the flf2v template
  uses SamplerEulerAncestral and the ic_lora template a plain KSampler with
  euler_ancestral + linear_quadratic. Lightricks' own example workflows use
  euler_ancestral_cfg_pp (single stage) and euler_cfg_pp (two stage).
  TI2VidTwoStagesHQPipeline swaps in the second-order res_2s sampler for
  fewer steps., resolution Generate low, then upscale. ComfyUI templates set
  EmptyLTXVLatentVideo to 768x512 / 97 frames and target 1280x720 (with a
  1920x1088 resize and a 1536-long-edge resize in the chain), using
  LTXVLatentUpsampler with ltx-2.3-spatial- upscaler-x2-1.1.safetensors.
  Lightricks' own example workflows use 960x544 / 121 frames with the same x2
  upscaler. Hard constraints from the model card: width and height must be
  divisible by 32, and frame count must be divisible by 8 plus 1 (97, 121,
  193...). Frame rate: LTXVConditioning is 24 in the t2v/id_lora templates
  and 25 in the ic_lora/flf2v templates; CreateVideo is 24 in the ComfyUI
  templates. The x2 temporal upscaler is described in the model card as being
  for higher FPS., lora_strength Official distilled/acceleration LoRA: 0.5 in
  every ComfyUI template (LoraLoaderModelOnly), 0.5 plus a 0.2 second branch
  in Lightricks' single- stage example workflow, and 0.8 in the ltx-pipelines
  README/installation CLI example (--distilled-lora <path> 0.8). Official IC-
  LoRA Union-Control and ID-LoRA: 1.0 in the templates. Community norm for
  content/style LoRAs: the IC-LoRA Dual-Character author states 0.6-1.0 used
  alone, dropped to 0.3-0.5 when stacked with other LoRAs..

───────────────────────────────────────────────────────────────────────────────
  ADOPTION & POPULARITY
───────────────────────────────────────────────────────────────────────────────
  Downloads               2,212
  Rating                  100% positive (150/150)
  Comments                17
  Published               173 days ago
  Download rate           7.7/day since release

  Adoption rank           #20 of 150 in this collection  (High)
  Momentum rank           #43 of 150 by download rate  (Steady)

  Adoption tier is the percentile of total downloads within this collection;
  momentum is the percentile of downloads-per-day since release. Momentum is
  ranked rather than measured against a fixed rate, because this base model
  is itself new and every LoRA looks fast on an absolute scale. Both numbers
  are CivitAI's, read on the scrape date at the foot of this file.

───────────────────────────────────────────────────────────────────────────────
  CHAINING / STACKING
───────────────────────────────────────────────────────────────────────────────
  Stacks with others      not stated

  The author gave no chaining guidance. The rules below are the
  general LTX-2.3 ones, not anything specific to this LoRA.

  General LTX-2.3 rules that apply here (see REFERENCE.txt):

  • Stacking an acceleration LoRA with a task LoRA is officially
    sanctioned: the ID-LoRA template chains LoraLoaderModelOnly(ltx_2.3_22b
    _distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16, 0.5) into
    LoraLoaderModelOnly(ltx-2.3-id-lora-talkvid-3k, 1.0) on the dev
    checkpoint.
  • The Python CLI exposes `--lora <path> [strength]` as a repeatable flag
    with default strength 1.0, so multiple adapters at once is a supported
    configuration, not a hack.
  • Load the acceleration/distilled LoRA on the dev checkpoint only. The
    two-stage pipelines require it for the full model but explicitly do not
    use it for DistilledPipeline, ICLoraPipeline or DubItPipeline, which
    already run distilled weights.
  • Do not stack two distillation LoRAs (e.g. the official rank-384 /
    rank-111 one plus a community DMD LoRA). The DaSiWa DMD LoRA already
    blends 75-80% of the official distilled delta into its audio-output
    route and is described by its author as recreating the distilled model
    on top of stock dev, i.e. a replacement.
  • IC-LoRAs are not chained like ordinary LoRAs: each carries a
    reference_downscale_factor in its safetensors metadata and needs a node
    that reads it (GetICLoRAParameters -> LTXVAddGuide in core, or
    LTXICLoRALoaderModelOnly + LTXAddVideoICLoRAGuide in ComfyUI-LTXVideo).
    Without one the factor silently defaults to 1.
  • Two IC-LoRA reference guides can be driven from a single IC-LoRA: the
    CrossView Warp author wires both a depth-warp video and the original
    video into separate reference guides and sets latent_downscale_factor =
    1 on both.
  • DubItPipeline accepts exactly one Dub-It IC-LoRA (`--lora` exactly
    once) and applies it in both stages — that path is single-adapter by
    design.
  • Community convention on strengths when stacking: the IC-LoRA Dual-
    Character author recommends 0.6-1.0 for a content LoRA used alone,
    dropped to 0.3-0.5 when combined with others.
  • Some LoRA pairs are designed to be blended against each other — the LTX
    2.3 Crisp Enhance / Soft Enhance pair is described by its author as
    "can be used together and balanced at different strengths".
  • The gemma-3-12b-it-abliterated LoRA is not a video LoRA and does not
    belong in the DiT chain — it patches the Gemma text encoder through a
    model+clip LoraLoader, as the official templates wire it.
  • The author never states which weight family this targets. Within
    LTX-2.3, dev vs distilled is low risk: they are the same architecture
    and the distillation itself is published as a LoRA that the model card
    calls "applicable to the full model", so ordinary content LoRAs load on
    both. Expect style/strength drift rather than breakage; the

───────────────────────────────────────────────────────────────────────────────
  PROS
───────────────────────────────────────────────────────────────────────────────
  + Well established — 2,212 downloads. [source: CivitAI stats]
  + Positively received — 150 thumbs-up with no down-votes. [source:
    CivitAI stats]
  + Explicit trigger word(s) — DISPSTYLE, char_rr, char_bb, char_case,
    char_invisi, char_prism, char_malev, char_roy, char_punchup — so the
    effect can be turned on and off from the prompt. [source: model
    metadata]

───────────────────────────────────────────────────────────────────────────────
  CONS & CAVEATS
───────────────────────────────────────────────────────────────────────────────
  - Author never states which generation mode it was trained for; test
    against your own workflow before relying on it. [source: description]
  - Large at 1286 MB — slower to load and heavier to stack. [source: file
    size]

───────────────────────────────────────────────────────────────────────────────
  COMPATIBILITY WARNING
───────────────────────────────────────────────────────────────────────────────
  Within LTX-2.3, dev vs distilled is low risk: they are the same
  architecture and the distillation itself is published as a LoRA that the
  model card calls "applicable to the full model", so ordinary content LoRAs
  load on both. Expect style/strength drift rather than breakage; the
  distilled LoRA is meant to be applied to a dev checkpoint, not stacked on
  an already-distilled one (the Python pipelines that start from a distilled
  checkpoint — DistilledPipeline, ICLoraPipeline, DubItPipeline — explicitly
  do not take --distilled-lora). The genuine risk is cross-version and cross-
  format, and it is silent rather than a hard error: the LTX-2 repo README
  states flatly that files "are not interchangeable between the two models,
  and a LoRA only works with the model it was trained on", and a wrong-
  version LoRA surfaces only as a stream of "lora key not loaded:
  diffusion_model...." console warnings while generation proceeds as if no
  LoRA were attached. The same silent no-op hits correctly-versioned LoRAs
  saved in a non-ComfyUI key layout: SimpleTuner issue #2349 shows keys like
  diffusion_model.connectors.audio_connector.transformer_blocks.N.attn1.to_q
  going unloaded, where the LTX-2 trainer's own layout is attn1/attn2 for
  video, audio_attn1/audio_attn2 for audio and audio_to_video_attn /
  video_to_audio_attn for cross-modal. Community finetunes (Sulphur 2, JoyAI-
  Echo merges) load stock LTX-2.3 LoRAs but shift the result. And an IC-LoRA
  loaded without a metadata-reading node loses its reference scale: ComfyUI's
  GetICLoRAParameters reads reference_downscale_factor out of the LoRA's
  safetensors metadata and silently falls back to 1 when it is absent, so a
  ref0.5 IC-LoRA wired straight into LTXVAddGuide without it feeds the
  reference at the wrong latent scale.

───────────────────────────────────────────────────────────────────────────────
  AUTHOR'S DESCRIPTION
───────────────────────────────────────────────────────────────────────────────
  This is just the same dataset but trained for LTX 2.3 but I increased rank
  from 32 to 64 (thus the 1.2GB file size). The result is really great. We
  get 6 character likeness and voices which sound and look great. For one
  lora that's a lot of information! And the characters don't bleed into
  eachother.
  Edit: All issues have been fixed in v1.1. Give it a try!
  Add "subtitle text" to negatives if you get them, you should not get them
  otherwise.
  Robert, Invisigal, Blonde Blazer (and without powers too), Prism, Malevola,
  Chase
  Punchup is trained but not enough data, haven't tested. (char_punchup)
  Royd is in the data as char_roy but also not much data.
  Other characters are in the dataset but not tagged. Shroud is not in the
  data at all.v1.0
  Style Trigger:
  DISPSTYLE
  Invisigal :
  char_invisi, a woman with short dark hair accented by a purple streak,
  wearing a purple jacket over a dark top and distressed black jeans
  Blonde Blazer:
  char_bb is a woman with long blonde hair and a blue mask, wearing a blue
  and yellow superhero suit with a yellow cape and a red diamond-shaped gem
  on her chest.
  (no costume, powerless)
  char_bb has long, wavy dark brown hair and wears a strapless evening dress
  with long dark blue opera gloves
  Robert Robertson:
  char_rr has short brown hair and wears a light blue button-down shirt with
  the sleeves rolled up to his elbows and dark trousers.
  Prism: char_prism, She is a black woman with straight hair split down the
  middle, teal on the viewer's left and magenta on the right. She wears
  large, rectangular, reflective teal visors, blue lipstick, and a gold ring
  necklace. Her attire consists of a black sleeveless top, thick gold bands
  on her upper arms, a long magenta glove on her left arm, and a teal glove
  on her right hand
  Malevola:
  char_malev a woman with red skin, long black horns, and glowing yellow
  eyes, a long red tail,, is dressed in a white tank top and denim shorts.
  she has a sword attached to her back.
  Chase:
  char_chase, an older Black man with white hair in locs and a mustache,
  wears a yellow sweater over a light blue collared shirt and black trousers
  with a gold buckle belt. A pair of glasses hangs from the collar of his
  sweater.

───────────────────────────────────────────────────────────────────────────────
  VERSION NOTES
───────────────────────────────────────────────────────────────────────────────
  Fixed the issue with the dataset causing visi and bb's voice to merge. I
  added in an additional 100 or so missing video clips of visi and around 30
  more of bb. Also I fixed some bad captions. Then I resumed from 24.5k steps
  and trained to 31k steps.
  The result is much better voices for the two. I missed a few captions on
  rr's beard so if you see stubble on female characters just put "stubble" in
  the negative prompt. Also roy and punchup's voice and likeness is slightly
  better now. But still not stable. Overall very good result. I'm really
  proud of this one.

───────────────────────────────────────────────────────────────────────────────
  EXAMPLE MEDIA
───────────────────────────────────────────────────────────────────────────────
  6 preview(s) in ./example_images/ — community results for
  this LoRA. Videos keep a matching _poster.jpg still frame.

  No prompts are recorded below: CivitAI's public API flags these posts as
  having generation metadata but does not return it, so the prompts cannot be
  scraped. Open the model page to read them.

  01_video.mp4  (video, 1920x1024)
      still: 01_poster.jpg
  02_video.mp4  (video, 1920x1024)
      still: 02_poster.jpg
  03_video.mp4  (video, 1920x1024)
      still: 03_poster.jpg
  04_video.mp4  (video, 1920x1024)
      still: 04_poster.jpg
  05_video.mp4  (video, 1920x1024)
      still: 05_poster.jpg
  06_video.mp4  (video, 1920x1024)
      still: 06_poster.jpg

───────────────────────────────────────────────────────────────────────────────
  HOW TO USE (COMFYUI)
───────────────────────────────────────────────────────────────────────────────
  1. Update ComfyUI. LTX-2.3 is supported in ComfyUI core
  (comfy_extras/nodes_lt.py provides EmptyLTXVLatentVideo, LTXVAddGuide,
  GetICLoRAParameters, LTXVConditioning, LTXVScheduler,
  LTXVConcatAVLatent/LTXVSeparateAVLatent, LTXVReferenceAudio and more) and
  the six official templates need no custom nodes; install the Lightricks
  "ComfyUI-LTXVideo" custom node pack only if you want IC-LoRA loaders, low-
  VRAM loaders, HDR decode, audio-only nodes or Q8 nodes.
  2. Put the base checkpoint in ComfyUI/models/checkpoints/ —
  ltx-2.3-22b-dev-fp8.safetensors (29.15 GB) or ltx-2.3-22b-distilled-
  fp8.safetensors (29.53 GB) from Lightricks/LTX-2.3-fp8, or the bf16
  originals (46.15 GB) from Lightricks/LTX-2.3. It is a single file that also
  carries the video VAE, the audio VAE and the text projection.
  3. Put gemma_3_12B_it_fp4_mixed.safetensors in
  ComfyUI/models/text_encoders/ (from Comfy-Org/ltx-2). The Python package
  instead wants the full google/gemma-3-12b-it-qat-q4_0-unquantized folder.
  4. Put ltx-2.3-spatial-upscaler-x2-1.1.safetensors in
  ComfyUI/models/latent_upscale_models/ (needed by every two-stage template),
  and moge_2_vitl_normal_fp16.safetensors (Comfy-Org/MoGe) in
  ComfyUI/models/geometry_estimation/ if you use the IC-LoRA depth path.
  5. Put every LoRA .safetensors in ComfyUI/models/loras/. Subfolders are
  fine — Lightricks' own workflows reference
  "ltxv/ltx2/ltx-2.3-22b-distilled-lora-384-1.1.safetensors".
  6. Load a normal content/style LoRA with LoraLoaderModelOnly, wired between
  CheckpointLoaderSimple's MODEL output and the guider/sampler. In the stock
  templates you chain it after the distilled-LoRA loader. Use the full
  LoraLoader (model + clip) only for the gemma-3-12b-it-abliterated LoRA,
  which patches the Gemma text encoder rather than the video model.
  7. On a dev checkpoint, also load the acceleration LoRA
  (ltx_2.3_22b_distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16.safetensors
  at 0.5) unless you intend to run the slow guided path.
  8. For an IC-LoRA, the LoRA loader alone is not enough: either use
  LTXICLoRALoaderModelOnly + LTXAddVideoICLoRAGuide (ComfyUI-LTXVideo), or —
  as the official ic_lora template does — feed the LoRA-loaded model through
  core's GetICLoRAParameters into LTXVAddGuide, so the
  reference_downscale_factor stored in the file's safetensors metadata is
  read instead of defaulting to 1.
  9. Template Library > Video > any LTX-2.3 workflow gives you correct wiring
  to start from; workflow JSONs are at Comfy-Org/workflow_templates/templates
  /video_ltx2_3_{t2v,i2v,ia2v,flf2v,ic_lora,id_lora}.json and
  Lightricks/ComfyUI-LTXVideo/example_workflows/2.3/.

───────────────────────────────────────────────────────────────────────────────
  Scraped by lorakit on 2026-09-04 23:14 UTC.
  Stats and description are the author's; pros/cons are derived from the
  evidence tagged beside each line.
───────────────────────────────────────────────────────────────────────────────