Add 🌸💀 DaSiWa LTX 2.3 DMD/Distilled LoRA 💀🌸

.gitattributes

@@ -351,3 +351,5 @@ muscle-growth/example_images/06_video.mp4 filter=lfs diff=lfs merge=lfs -text
 ltx2-3-aesthetic-of-asian-portraits/example_images/01_poster.jpg filter=lfs diff=lfs merge=lfs -text
 ltx2-3-aesthetic-of-asian-portraits/example_images/01_video.mp4 filter=lfs diff=lfs merge=lfs -text
 ltx2-3-aesthetic-of-asian-portraits/example_images/02_video.mp4 filter=lfs diff=lfs merge=lfs -text
+dasiwa-ltx-2-3-dmd-distilled-lora/example_images/01_poster.jpg filter=lfs diff=lfs merge=lfs -text
+dasiwa-ltx-2-3-dmd-distilled-lora/example_images/01_video.mp4 filter=lfs diff=lfs merge=lfs -text

dasiwa-ltx-2-3-dmd-distilled-lora/LTX2.3_Hybrid_DMD_Distill_global-r72_standard.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a69ebdf1bdaadcbbe9ce51d919dfe10fda0ea718bef44ce4fb02dc6f4dae8f
+size 1210688080

dasiwa-ltx-2-3-dmd-distilled-lora/example_images/01_video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f0a74540a0d3086d8ade6e45a10ff8b064852413b6fd05a298e4b03148c419b
+size 1799192

dasiwa-ltx-2-3-dmd-distilled-lora/info.txt

@@ -0,0 +1,278 @@
+═══════════════════════════════════════════════════════════════════════════════
+  🌸💀 DASIWA LTX 2.3 DMD/DISTILLED LORA 💀🌸
+═══════════════════════════════════════════════════════════════════════════════
+
+───────────────────────────────────────────────────────────────────────────────
+  IDENTITY
+───────────────────────────────────────────────────────────────────────────────
+  Base model        LTX-2.3
+  Generation mode   Image-to-Video (I2V)
+  Checkpoint        ltx-2.3-22b-dev (full / base)
+  Category          Speed / step-reduction
+  Author            Darksidewalker
+  Source            https://civitai.com/models/2783959
+  Provenance        CivitAI · model 2783959 · version 3135769
+
+───────────────────────────────────────────────────────────────────────────────
+  WEIGHTS
+───────────────────────────────────────────────────────────────────────────────
+  File              LTX2.3_Hybrid_DMD_Distill_global-r72_standard.safetensors
+  Version           r72-L75-D25-v1.0
+  Size              1,154.6 MB
+  Format            SafeTensor / unknown precision
+  SHA256            a2a69ebdf1bdaadcbbe9ce51d919dfe10fda0ea718bef44ce4fb02dc6f4dae8f
+  Published         2026-07-16
+
+───────────────────────────────────────────────────────────────────────────────
+  TRIGGER WORDS
+───────────────────────────────────────────────────────────────────────────────
+  No trigger word. The LoRA applies as soon as it is loaded —
+  strength is the only way to dial the effect up or down.
+
+───────────────────────────────────────────────────────────────────────────────
+  RECOMMENDED SETTINGS
+───────────────────────────────────────────────────────────────────────────────
+  Taken verbatim from the author's description:
+
+  LoRA strength       0.4–1.0
+  Sampler steps       8–12
+
+───────────────────────────────────────────────────────────────────────────────
+  ADOPTION & POPULARITY
+───────────────────────────────────────────────────────────────────────────────
+  Downloads               1,595
+  Rating                  100% positive (62/62)
+  Comments                0
+  Published               50 days ago
+  Download rate           7.9/day since release
+
+  Adoption rank           #30 of 150 in this collection  (High)
+  Momentum rank           #42 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 step-reduction adapter for LTX-2.3 is first-party. Lightricks
+    publishes ltx-2.3-22b-distilled-lora-384.safetensors and
+    ltx-2.3-22b-distilled-lora-384-1.1.safetensors (rank 384, 7.61 GB each,
+    verified from the HF file listing), described in the model card as "A
+    LoRA version of the distilled model applicable to the full model" —
+    apply it to ltx-2.3-22b-dev and the model runs the distilled 8-step,
+    CFG=1 schedule. Comfy-Org additionally ships a dynamically rank-reduced
+    repack, ltx_2.3_22b_distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16.
+    safetensors (2.74 GB), and this is the file every offic
+  • The full, flexible, trainable bf16 model — ltx-2.3-22b-dev.safetensors
+    is 46.15 GB bf16, 29.15 GB as ltx-2.3-22b-dev-fp8.safetensors. This is
+    what LoRAs are trained against, and what four of the six official
+    ComfyUI LTX-2.3 templates load (t2v, i2v, ia2v, id_lora). Needs either
+    real guidance (Lightricks' own single-stage workflow sets LTXVScheduler
+    to 15 steps with MultimodalGuider) or the distill
+
+───────────────────────────────────────────────────────────────────────────────
+  PROS
+───────────────────────────────────────────────────────────────────────────────
+  + Well established — 1,595 downloads. [source: CivitAI stats]
+  + Positively received — 62 thumbs-up with no down-votes. [source: CivitAI
+    stats]
+  + Author published concrete recommended settings. [source: description]
+
+───────────────────────────────────────────────────────────────────────────────
+  CONS & CAVEATS
+───────────────────────────────────────────────────────────────────────────────
+  - Author labels this experimental — results vary shot to shot. [source:
+    description]
+  - No trigger word — the LoRA is always on once loaded, so strength is the
+    only control. [source: model metadata]
+  - Large at 1155 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 an experimental alternate distillation LoRA based on JoyEcho DMD
+  for LTX 2.3.
+
+  Tested settings:
+  - Strength 0.4-1.0
+
+  - Works with: 1st pass and 2nd pass and 3rd pass
+
+  - 8-12 steps
+
+  Limited testing was done. Good luck! ✌️
+  The Goal:
+  This LoRA recreates the distilled JoyEcho LTX 2.3 model changes on top of
+  the original LTX 2.3 Dev checkpoint, so it keeps JoyEcho’s enhanced
+  distilled behavior.
+
+  All standard model routes are included. Unlike an I2V CondSafe LoRA,
+  nothing is intentionally removed.
+  V2
+  I raised the overall rank and did a selective SVD to get more quality.
+  Sound and artifacts are noticeable better on this one, from my testing. The
+  final audio-output route is blended: 20% of the JoyEcho change and 80% of
+  the official LTX distilled LoRA change.
+  The LoRA is compressed dynamically to a maximum rank of 86 per layer while
+  targeting 99% retained weight energy. And target 160 rank for Audio.
+  V1
+  The final audio-output route is blended: 25% of the JoyEcho change and 75%
+  of the official LTX distilled LoRA change. This keeps the better distilled
+  audio behavior from the official r384 while reducing the overly loud sound
+  produced by the full JoyEcho output route.
+
+  The LoRA is compressed dynamically to a maximum rank of 72 per layer while
+  targeting 96% retained weight energy. This makes it substantially smaller
+  and more practical than a full checkpoint delta, while preserving the most
+  important learned changes.
+
+───────────────────────────────────────────────────────────────────────────────
+  VERSION NOTES
+───────────────────────────────────────────────────────────────────────────────
+  V1
+  - DMD x LTX blend
+
+  - audio-output route is blended: 25% of the JoyEcho and 75% of the official
+  LTX
+
+───────────────────────────────────────────────────────────────────────────────
+  EXAMPLE MEDIA
+───────────────────────────────────────────────────────────────────────────────
+  1 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, 1088x1408)
+      still: 01_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 23:16 UTC.
+  Stats and description are the author's; pros/cons are derived from the
+  evidence tagged beside each line.
+───────────────────────────────────────────────────────────────────────────────

dasiwa-ltx-2-3-dmd-distilled-lora/metadata.json

@@ -0,0 +1,112 @@
+{
+  "name": "🌸💀 DaSiWa LTX 2.3 DMD/Distilled LoRA 💀🌸",
+  "slug": "dasiwa-ltx-2-3-dmd-distilled-lora",
+  "base_model": "LTX-2.3",
+  "media_type": "video",
+  "model_id": 2783959,
+  "version_id": 3135769,
+  "version_name": "r72-L75-D25-v1.0",
+  "source": "civitai",
+  "url": "https://civitai.com/models/2783959",
+  "repo_id": null,
+  "file_path": null,
+  "author": "Darksidewalker",
+  "category": {
+    "key": "accelerator",
+    "label": "Speed / step-reduction"
+  },
+  "modalities": [
+    "i2v"
+  ],
+  "primary_modality": "i2v",
+  "checkpoint_family": "dev",
+  "trigger_words": [],
+  "recommended_settings": {
+    "LoRA strength": "0.4–1.0",
+    "Sampler steps": "8–12"
+  },
+  "chaining": {
+    "author": {
+      "stacks": "not stated",
+      "notes": "",
+      "avoid": [],
+      "requires": [],
+      "order": "",
+      "strength": "",
+      "stated": false
+    },
+    "general_rules": [
+      "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 step-reduction adapter for LTX-2.3 is first-party. Lightricks publishes ltx-2.3-22b-distilled-lora-384.safetensors and ltx-2.3-22b-distilled-lora-384-1.1.safetensors (rank 384, 7.61 GB each, verified from the HF file listing), described in the model card as \"A LoRA version of the distilled model applicable to the full model\" — apply it to ltx-2.3-22b-dev and the model runs the distilled 8-step, CFG=1 schedule. Comfy-Org additionally ships a dynamically rank-reduced repack, ltx_2.3_22b_distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16.safetensors (2.74 GB), and this is the file every offic",
+      "The full, flexible, trainable bf16 model — ltx-2.3-22b-dev.safetensors is 46.15 GB bf16, 29.15 GB as ltx-2.3-22b-dev-fp8.safetensors. This is what LoRAs are trained against, and what four of the six official ComfyUI LTX-2.3 templates load (t2v, i2v, ia2v, id_lora). Needs either real guidance (Lightricks' own single-stage workflow sets LTXVScheduler to 15 steps with MultimodalGuider) or the distill"
+    ]
+  },
+  "pros": [
+    "Well established — 1,595 downloads. [source: CivitAI stats]",
+    "Positively received — 62 thumbs-up with no down-votes. [source: CivitAI stats]",
+    "Author published concrete recommended settings. [source: description]"
+  ],
+  "cons": [
+    "Author labels this experimental — results vary shot to shot. [source: description]",
+    "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]",
+    "Large at 1155 MB — slower to load and heavier to stack. [source: file size]"
+  ],
+  "file": {
+    "name": "LTX2.3_Hybrid_DMD_Distill_global-r72_standard.safetensors",
+    "size_bytes": 1210688080,
+    "size_mb": 1154.6,
+    "sha256": "a2a69ebdf1bdaadcbbe9ce51d919dfe10fda0ea718bef44ce4fb02dc6f4dae8f"
+  },
+  "stats": {
+    "downloads": 1595,
+    "thumbs_up": 62,
+    "thumbs_down": 0
+  },
+  "popularity": {
+    "downloads": 1595,
+    "thumbs_up": 62,
+    "thumbs_down": 0,
+    "comments": 0,
+    "approval": "100% positive (62/62)",
+    "age_days": 50.2,
+    "downloads_per_day": 7.89,
+    "rank_by_downloads": 30,
+    "rank_by_momentum": 42,
+    "of_total": 150,
+    "adoption_tier": "High",
+    "momentum": "Steady"
+  },
+  "selection_score": 2339.0,
+  "selection_reasons": [
+    "1,595 downloads",
+    "62 thumbs-up",
+    "top-ranked for its category"
+  ],
+  "example_media": [
+    {
+      "file": "01_video.mp4",
+      "type": "video",
+      "resolution": "1088x1408",
+      "prompt": "",
+      "poster": "01_poster.jpg"
+    }
+  ],
+  "tags": [
+    "ltx2.3",
+    "lightspeed",
+    "dmd",
+    "distillation",
+    "base model"
+  ],
+  "scraped_at": "2026-09-04T23:16:48+00:00",
+  "path": "ltxv-2.3/dasiwa-ltx-2-3-dmd-distilled-lora"
+}