Instructions to use Alissonerdx/EditAnything with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use Alissonerdx/EditAnything with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("Lightricks/LTX-2.3", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("Alissonerdx/EditAnything") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
Add Edit Anything v2 (LTX-2.5) section: usage, prompt format, training
Browse files
README.md
CHANGED
|
@@ -1,539 +1,660 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
library_name: diffusers
|
| 4 |
-
base_model:
|
| 5 |
-
|
| 6 |
-
-
|
| 7 |
-
|
| 8 |
-
-
|
| 9 |
-
-
|
| 10 |
-
--
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
---
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
-
|
| 37 |
-
|
| 38 |
-
##
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
###
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
the
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
###
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
*
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
-
|
| 309 |
-
|
| 310 |
-
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
|
| 316 |
-
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
- **
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
|
| 363 |
-
|
| 364 |
-
|
| 365 |
-
|
| 366 |
-
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
|
| 372 |
-
|
| 373 |
-
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
-
|
| 385 |
-
|
| 386 |
-
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
|
| 392 |
-
|
| 393 |
-
|
| 394 |
-
|
| 395 |
-
|
| 396 |
-
|
| 397 |
-
|
| 398 |
-
|
| 399 |
-
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
|
| 404 |
-
|
| 405 |
-
|
| 406 |
-
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
- *
|
| 415 |
-
|
| 416 |
-
- *
|
| 417 |
-
|
| 418 |
-
- *
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
|
| 422 |
-
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
|
| 429 |
-
|
| 430 |
-
|
| 431 |
-
|
| 432 |
-
|
| 433 |
-
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
| 439 |
-
|
| 440 |
-
|
| 441 |
-
|
| 442 |
-
|
| 443 |
-
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
|
| 451 |
-
-
|
| 452 |
-
|
| 453 |
-
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
|
| 457 |
-
|
| 458 |
-
|
| 459 |
-
|
| 460 |
-
|
| 461 |
-
|
| 462 |
-
|
| 463 |
-
|
| 464 |
-
|
| 465 |
-
|
| 466 |
-
|
| 467 |
-
|
| 468 |
-
-
|
| 469 |
-
|
| 470 |
-
|
| 471 |
-
|
| 472 |
-
|
| 473 |
-
**
|
| 474 |
-
|
| 475 |
-
|
| 476 |
-
|
| 477 |
-
|
| 478 |
-
|
| 479 |
-
|
| 480 |
-
|
| 481 |
-
|
| 482 |
-
|
| 483 |
-
|
| 484 |
-
|
| 485 |
-
|
| 486 |
-
|
| 487 |
-
|
| 488 |
-
|
| 489 |
-
|
| 490 |
-
|
| 491 |
-
|
| 492 |
-
|
| 493 |
-
|
| 494 |
-
|
| 495 |
-
|
| 496 |
-
|
| 497 |
-
|
| 498 |
-
|
| 499 |
-
|
| 500 |
-
|
| 501 |
-
|
| 502 |
-
|
| 503 |
-
|
| 504 |
-
|
| 505 |
-
|
| 506 |
-
|
| 507 |
-
|
| 508 |
-
|
| 509 |
-
|
| 510 |
-
-
|
| 511 |
-
|
| 512 |
-
|
| 513 |
-
|
| 514 |
-
|
| 515 |
-
-
|
| 516 |
-
|
| 517 |
-
|
| 518 |
-
|
| 519 |
-
|
| 520 |
-
|
| 521 |
-
|
| 522 |
-
|
| 523 |
-
|
| 524 |
-
|
| 525 |
-
|
| 526 |
-
|
| 527 |
-
|
| 528 |
-
|
| 529 |
-
|
| 530 |
-
|
| 531 |
-
|
| 532 |
-
|
| 533 |
-
|
| 534 |
-
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
|
| 538 |
-
|
| 539 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: diffusers
|
| 4 |
+
base_model:
|
| 5 |
+
- Lightricks/LTX-2.3
|
| 6 |
+
- Lightricks/LTX-2.5
|
| 7 |
+
tags:
|
| 8 |
+
- lora
|
| 9 |
+
- video
|
| 10 |
+
- video-editing
|
| 11 |
+
- ltx-2.3
|
| 12 |
+
- ltx-2.5
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Edit Anything β Experimental LTX-2 Video Editing LoRAs
|
| 16 |
+
|
| 17 |
+
> **Heads up.** These LoRAs are research experiments. They are far from
|
| 18 |
+
> production-ready and will fail on many inputs. They are released for the
|
| 19 |
+
> community to play with and break, not as a finished tool.
|
| 20 |
+
|
| 21 |
+
This repository hosts four unrelated training tracks:
|
| 22 |
+
|
| 23 |
+
0. **Edit Anything v2 for LTX-2.5** β the current one, a plain instruction-editing
|
| 24 |
+
IC-LoRA. Start here. *(built on LTX-2.5 22B)*
|
| 25 |
+
1. **Edit Anything v0.1 β motion transfer LoRA** (two ranks). *(LTX-2.3)*
|
| 26 |
+
2. **Edit Anything β no-reference multitask LoRA** (rank 256, prompt-driven only). *(LTX-2.3)*
|
| 27 |
+
3. **Reference video-to-video (Ref V2V) β experimental IC-LoRA + sidecar modules** (two builds). *(LTX-2.3)*
|
| 28 |
+
|
| 29 |
+
Tracks 1-3 are **LTX-2.3** and are kept for reference. Track 0 is **LTX-2.5** and is
|
| 30 |
+
the one being worked on.
|
| 31 |
+
|
| 32 |
+
Inference is meant to run through the **BFSnodes** ComfyUI custom nodes β
|
| 33 |
+
the Ref V2V build in particular needs them to load the sidecar modules and
|
| 34 |
+
install the custom branches into the transformer.
|
| 35 |
+
|
| 36 |
+
---
|
| 37 |
+
|
| 38 |
+
## 0. Edit Anything v2 β LTX-2.5
|
| 39 |
+
|
| 40 |
+
File: `edit_anything_v2_ltx2.5.safetensors` (rank 128)
|
| 41 |
+
Workflow: `workflows/EditAnythingLTX2.5.json`
|
| 42 |
+
Base model: **LTX-2.5 22B dev**
|
| 43 |
+
|
| 44 |
+
### What it is
|
| 45 |
+
|
| 46 |
+
In this experimental version it is possible to perform several operations together in a
|
| 47 |
+
single prompt β **Replace, Remove, Add, Restyle** β and everything becomes easier with the
|
| 48 |
+
help of the prompt enhancer, so you can write in natural language and it converts it to the
|
| 49 |
+
appropriate prompt.
|
| 50 |
+
|
| 51 |
+
Unlike track 3, this LoRA takes **no reference image**. It is a plain IC-LoRA: your source
|
| 52 |
+
video goes in as the guide, the instruction goes in as text, and the edited video comes out.
|
| 53 |
+
No RoPE tricks, no sidecar modules, no loss masks β the simplest thing that works.
|
| 54 |
+
|
| 55 |
+
### How to use it in ComfyUI
|
| 56 |
+
|
| 57 |
+
Load `workflows/EditAnythingLTX2.5.json`. If you are wiring it by hand with the **BFSnodes**
|
| 58 |
+
`LTX Multiple Controls` node, the settings that match training are:
|
| 59 |
+
|
| 60 |
+
| input | value |
|
| 61 |
+
|---|---|
|
| 62 |
+
| `guide_video` | your source video |
|
| 63 |
+
| `guide_source_id` | `0` |
|
| 64 |
+
| `guide_layout` | `overlap` |
|
| 65 |
+
| `guide_ref_resize_mode` | `match_target` |
|
| 66 |
+
| `guide_downscale_factor` | `1` |
|
| 67 |
+
|
| 68 |
+
Leave `identity_image`, `mask_video` and `identity_mask_image` **disconnected** β this LoRA
|
| 69 |
+
was trained without any of them.
|
| 70 |
+
|
| 71 |
+
Trained resolution is **704x384, 73 frames**. Other sizes and lengths are extrapolation and
|
| 72 |
+
degrade first in temporal consistency. Suggested start: LoRA strength `1.0`, CFG `3-5`,
|
| 73 |
+
30 steps, no LightX2V.
|
| 74 |
+
|
| 75 |
+
### Prompting
|
| 76 |
+
|
| 77 |
+
Every prompt starts with the trigger:
|
| 78 |
+
|
| 79 |
+
```
|
| 80 |
+
edit_anything:
|
| 81 |
+
```
|
| 82 |
+
|
| 83 |
+
Then one sentence per edit. The four verbs below are the only ones the model was trained on;
|
| 84 |
+
the percentage is that verb's share of the 6829 training instructions:
|
| 85 |
+
|
| 86 |
+
```
|
| 87 |
+
Replace <the existing thing> with <the new thing, described>. 50%
|
| 88 |
+
Add a <object with attributes> <where>. 26%
|
| 89 |
+
Remove <the thing, identified precisely>. 18%
|
| 90 |
+
Restyle <the region, identified> into <style> style. 5%
|
| 91 |
+
```
|
| 92 |
+
|
| 93 |
+
Rules that come straight from the training distribution β breaking them is the usual reason
|
| 94 |
+
a prompt does nothing:
|
| 95 |
+
|
| 96 |
+
- **10 to 20 words per instruction** (median 15). Shorter loses the description the model
|
| 97 |
+
needs; longer drifts off-distribution.
|
| 98 |
+
- **One object per `Add`.** 95% of trained Add instructions begin `Add a ...` / `Add an ...`.
|
| 99 |
+
For a group or a crowd use `Replace` on an existing region instead of `Add several ...`.
|
| 100 |
+
- **Anchor each instruction to something visible**, the way the data does: "to the left of
|
| 101 |
+
the man", "on the desk on the right side of the frame", "behind the speaker". 63% of the
|
| 102 |
+
trained Add instructions carry an anchor like this.
|
| 103 |
+
- **Describe state, not manner.** "standing in the green grass field", "sitting upright on
|
| 104 |
+
the desk" β not "swimming gracefully".
|
| 105 |
+
- **`Restyle` targets a region, not the whole clip.** 93% of trained Restyle instructions
|
| 106 |
+
name a wall, a screen, a background or a face. A whole-video restyle is off-distribution;
|
| 107 |
+
express it as two or three Restyle instructions over the regions that carry the look.
|
| 108 |
+
- **`Make it <style>` and `Turn it into <style>` do not exist** in this model's training
|
| 109 |
+
data β zero occurrences. Use `Restyle`.
|
| 110 |
+
|
| 111 |
+
Real examples, in the exact shape the model learned:
|
| 112 |
+
|
| 113 |
+
```
|
| 114 |
+
edit_anything: Add a medium-sized brown dog sitting on the grass to the right of the
|
| 115 |
+
motorcycles. Remove the black jacket worn by the smiling man in the left foreground.
|
| 116 |
+
Replace the man's graying hair with bright red hair.
|
| 117 |
+
|
| 118 |
+
edit_anything: Replace the gray and black hoodie worn by the man with a bright red zip-up
|
| 119 |
+
jacket with white stripes on the sleeves.
|
| 120 |
+
|
| 121 |
+
edit_anything: Remove the smartphone held in the man's hand.
|
| 122 |
+
|
| 123 |
+
edit_anything: Restyle the large blue digital screen behind the man into flat vector style.
|
| 124 |
+
```
|
| 125 |
+
|
| 126 |
+
Write one instruction per edit you actually want. Do not pad the prompt with edits you did
|
| 127 |
+
not ask for β an unrequested sky or background change will show up in the output.
|
| 128 |
+
|
| 129 |
+
### Training
|
| 130 |
+
|
| 131 |
+
- **Base**: LTX-2.5 22B dev, trained from scratch (no continuation from the 2.3 LoRAs).
|
| 132 |
+
- **Data**: 2750 aligned source/edited video pairs from **CoinVE-200K**, at 704x384x73.
|
| 133 |
+
- **Recipe**: IC-LoRA with a single conditioning stream (the source video, `overlap`,
|
| 134 |
+
source id 0). Rank 128 / alpha 128, AdamW `1e-4` constant, batch 1, 10 000 steps.
|
| 135 |
+
- **This release is a merge**: the weights are the 50/50 average of the step-7000 and
|
| 136 |
+
step-9500 checkpoints, reduced back to rank 128 by truncated SVD (Frobenius-optimal;
|
| 137 |
+
99.99% of the energy preserved). The earlier checkpoint generalises better outside the
|
| 138 |
+
training distribution, the later one follows instructions better inside it β the average
|
| 139 |
+
beat both parents on 4 of the 5 metrics measured on held-out data.
|
| 140 |
+
|
| 141 |
+
### Limitations
|
| 142 |
+
|
| 143 |
+
- Trained only on CoinVE-style edits. Requests far from that vocabulary get weaker results.
|
| 144 |
+
- Edits that change very little of the frame (subtle removals) are the least reliable.
|
| 145 |
+
- No reference image support β you cannot say "add *this* object" with a picture. That is
|
| 146 |
+
what track 3 tried, and it is still unsolved.
|
| 147 |
+
- 704x384x73 is the trained box; longer clips lose temporal consistency first.
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
---
|
| 151 |
+
|
| 152 |
+
## 1. Edit Anything v0.1 (motion transfer)
|
| 153 |
+
|
| 154 |
+
Files:
|
| 155 |
+
|
| 156 |
+
- `edit_anything_30k_v0.1_motion_transfer_r128.safetensors`
|
| 157 |
+
- `edit_anything_30k_v0.1_motion_transfer_r256.safetensors`
|
| 158 |
+
|
| 159 |
+
### What it is
|
| 160 |
+
|
| 161 |
+
**v0.1 is not a direct continuation of v1.0.** It was trained from scratch
|
| 162 |
+
in two stages:
|
| 163 |
+
|
| 164 |
+
1. **Stage 1 β image-only pretraining.** ~30 000 image edit pairs. Training
|
| 165 |
+
a *video* model on still images is admittedly not ideal, but it was a way
|
| 166 |
+
to push the editing vocabulary beyond what a small video-only dataset can
|
| 167 |
+
teach.
|
| 168 |
+
2. **Stage 2 β video fine-tune with `first_frame_conditioning > 0`.** This
|
| 169 |
+
restored the temporal prior and unlocked the motion-transfer behaviour
|
| 170 |
+
described below.
|
| 171 |
+
|
| 172 |
+
In theory v0.1 can do the same edits as v1.0, but **temporal consistency may
|
| 173 |
+
be weaker than v1.0** because so much of stage 1 happened on still images.
|
| 174 |
+
Test against v1.0 case-by-case before assuming v0.1 wins on your task.
|
| 175 |
+
|
| 176 |
+
### Motion transfer
|
| 177 |
+
|
| 178 |
+
Because stage 2 included first-frame conditioning, you can drive the LoRA
|
| 179 |
+
into a motion-transfer mode:
|
| 180 |
+
|
| 181 |
+
1. Take a guide video.
|
| 182 |
+
2. **Replace its first frame** with an edited still (insert a new subject,
|
| 183 |
+
swap an object, etc.). Use a strong image-editing model β **Flux Klein**
|
| 184 |
+
or similar β to prepare it; the quality of this single frame propagates
|
| 185 |
+
through the whole clip.
|
| 186 |
+
3. Feed the edited frame as the first frame of the input, and the original
|
| 187 |
+
guide video as the motion source.
|
| 188 |
+
|
| 189 |
+
The model uses the new first frame as the appearance anchor and copies the
|
| 190 |
+
motion from the rest of the guide.
|
| 191 |
+
|
| 192 |
+
Limitations (these are real, not theoretical β expect them to bite):
|
| 193 |
+
|
| 194 |
+
- **Hard scene cuts break it.** The model assumes continuous motion from
|
| 195 |
+
the first frame onwards. A cut to a different camera angle or location
|
| 196 |
+
mid-clip will produce smearing, ghosting, or the inserted subject jumping
|
| 197 |
+
to the wrong position. Use clips without cuts, or split at the cuts and
|
| 198 |
+
process each segment separately.
|
| 199 |
+
- **Very fast motion fails.** Quick pans, fast subject movement, or
|
| 200 |
+
high-velocity action confuse the motion-copy mechanism. Outputs degrade
|
| 201 |
+
to blur or to the model "freezing" on the first-frame appearance and
|
| 202 |
+
losing the motion entirely. Stick to moderate-speed clips.
|
| 203 |
+
- Poor blending / artefacts in the first frame propagate everywhere.
|
| 204 |
+
- Works best when the inserted subject roughly occupies the same region as
|
| 205 |
+
whatever it replaces.
|
| 206 |
+
|
| 207 |
+
### Prompting
|
| 208 |
+
|
| 209 |
+
Prompt is just as critical as in v1.0. **Describe both the object being
|
| 210 |
+
replaced and the new one in detail**. Example: *"Replace the bronze statue on
|
| 211 |
+
the left with a tall man wearing a navy raincoat and brown boots."* Vague
|
| 212 |
+
prompts produce bad edits.
|
| 213 |
+
|
| 214 |
+
### Which rank to use
|
| 215 |
+
|
| 216 |
+
The same training produced both files. v0.1 is actually the merge of the
|
| 217 |
+
two-stage training (one LoRA per stage), re-extracted at two different ranks
|
| 218 |
+
via Frobenius-optimal truncated SVD:
|
| 219 |
+
|
| 220 |
+
| File | Rank | Size | Frobenius retention |
|
| 221 |
+
|---|---|---|---|
|
| 222 |
+
| `edit_anything_30k_v0.1_motion_transfer_r128.safetensors` | 128 | 1.31 GB | ~99.4% |
|
| 223 |
+
| `edit_anything_30k_v0.1_motion_transfer_r256.safetensors` | 256 | 2.62 GB | ~99.9% |
|
| 224 |
+
|
| 225 |
+
r256 is closer to the merged source. r128 is normally indistinguishable in
|
| 226 |
+
practice. Pick whichever fits your workflow.
|
| 227 |
+
|
| 228 |
+
### How to wire the LoopingSampler
|
| 229 |
+
|
| 230 |
+
This is a **standard LoRA**, not a sidecar. Load it through the regular
|
| 231 |
+
ComfyUI LoraLoader **before** the LoopingSampler. On the sampler itself:
|
| 232 |
+
|
| 233 |
+
- `editanything_module` β **leave disconnected**.
|
| 234 |
+
- `ref_image` β the edited first frame (for motion transfer) **or** the
|
| 235 |
+
source frame you want preserved (for plain editing).
|
| 236 |
+
- `guide_frames` β the guide video.
|
| 237 |
+
- `enable_role_embedding`, `enable_adaln`, `enable_visual_crossattn` β
|
| 238 |
+
all **off**. None of those branches were trained for v0.1; turning them
|
| 239 |
+
on with no module connected does nothing anyway, but keeping them off
|
| 240 |
+
silences the WARN logs.
|
| 241 |
+
|
| 242 |
+
---
|
| 243 |
+
|
| 244 |
+
## 2. Edit Anything β no-reference multitask LoRA
|
| 245 |
+
|
| 246 |
+
File:
|
| 247 |
+
|
| 248 |
+
- `edit_anything_v1.1_r256.safetensors`
|
| 249 |
+
|
| 250 |
+
### What it is
|
| 251 |
+
|
| 252 |
+
A **prompt-only** multitask editing LoRA. No reference image, no first-frame
|
| 253 |
+
conditioning β the model is driven entirely by the text prompt and the
|
| 254 |
+
guide video. Trained on a balanced mix of **Add, Remove, Replace, Style**
|
| 255 |
+
edits.
|
| 256 |
+
|
| 257 |
+
### What it's different about it (vs v0.1)
|
| 258 |
+
|
| 259 |
+
The task vocabulary overlaps heavily with v0.1 β both can do Add, Remove,
|
| 260 |
+
Replace, Change, Convert. What changes here:
|
| 261 |
+
|
| 262 |
+
- **Two-stage training continuation**: the first stage gave the model its
|
| 263 |
+
edit vocabulary; the second stage refined it on a larger, more balanced
|
| 264 |
+
video pair set covering Add / Remove / Replace / Style.
|
| 265 |
+
- **Rank 256** (vs v0.1's effective rank from the merge), giving more
|
| 266 |
+
capacity for the broader task mix.
|
| 267 |
+
- Trained directly on video pairs, so the temporal behaviour on these
|
| 268 |
+
tasks tends to be steadier than on a model whose first stage was on
|
| 269 |
+
still images.
|
| 270 |
+
|
| 271 |
+
### How to use it
|
| 272 |
+
|
| 273 |
+
**Standalone** β load it as a regular LoRA on vanilla LTX-2.3 through any
|
| 274 |
+
ComfyUI LoRA loader. The file already carries everything it needs; no
|
| 275 |
+
stacking with v0.1, no companion module.
|
| 276 |
+
|
| 277 |
+
### Limitations
|
| 278 |
+
|
| 279 |
+
- No reference image β identity is not anchored, so Add / Replace of a
|
| 280 |
+
specific person or object will be wobblier than the Ref V2V build.
|
| 281 |
+
- No motion transfer (that's v0.1 only).
|
| 282 |
+
|
| 283 |
+
### Prompting
|
| 284 |
+
|
| 285 |
+
Same imperative shape as v0.1, but the training set split into four very
|
| 286 |
+
distinct caption styles. Match the one that fits the edit you want β the
|
| 287 |
+
distribution is narrow and the model expects the right shape.
|
| 288 |
+
|
| 289 |
+
The training set is roughly balanced across **Add, Remove, Replace and
|
| 290 |
+
Style** buckets, with Style being the smallest of the four. Captions
|
| 291 |
+
below are real examples drawn from those buckets.
|
| 292 |
+
|
| 293 |
+
#### Add β 15 to 30+ words, describe what to add and where
|
| 294 |
+
|
| 295 |
+
* `Add a smiling woman with brown hair, wearing a pink sleeveless top, sitting to the right of the man at the news desk.`
|
| 296 |
+
* `Add a person wearing a blue denim shirt over a white t-shirt to the right side of the frame, behind the person cooking.`
|
| 297 |
+
* `Add a decorated Christmas tree with red and white ornaments and lights to the right of the man.`
|
| 298 |
+
* `Add a blonde boy wearing a black t-shirt with a blue collar and blue patterned pants, sitting behind the other children in the upper center of the frame.`
|
| 299 |
+
* `Add two horizontal wooden strips to the front of the white range hood.`
|
| 300 |
+
|
| 301 |
+
Pattern: `Add <detailed subject description>, <position in frame>, <surrounding context>.`
|
| 302 |
+
|
| 303 |
+
#### Remove β very short, 4 to 10 words
|
| 304 |
+
|
| 305 |
+
* `Remove the man drinking from a glass.`
|
| 306 |
+
* `Remove the disco ball.`
|
| 307 |
+
* `Remove the large tree on the right.`
|
| 308 |
+
* `Remove the squirrel in the foreground.`
|
| 309 |
+
* `Remove the man on the left.`
|
| 310 |
+
|
| 311 |
+
Pattern: `Remove the <object>` (+ optional position). Resist the urge to
|
| 312 |
+
over-describe β long Remove prompts drift outside the training shape and
|
| 313 |
+
often fail.
|
| 314 |
+
|
| 315 |
+
#### Replace β 20 to 35 words, describe both old and new
|
| 316 |
+
|
| 317 |
+
* `Replace the white panel door on the right side of the frame with a dark brown grandfather clock.`
|
| 318 |
+
* `Replace the light-colored cat lying on the mat on the floor with a young woman sitting on the mat.`
|
| 319 |
+
* `Replace the dark grey knitted sweater on the man's torso with a black and white patterned Christmas sweater.`
|
| 320 |
+
* `Replace the blue robot with a glowing blue face on the left with a smiling man wearing sunglasses and a blue shirt.`
|
| 321 |
+
* `Replace the sitting person wearing a black cape on the left with a black fabric draped over an object.`
|
| 322 |
+
|
| 323 |
+
Pattern: `Replace <description of the original subject and its location> with <description of the new subject>.`
|
| 324 |
+
|
| 325 |
+
#### Style β fixed template, the style name is what changes
|
| 326 |
+
|
| 327 |
+
* `Convert the video into a Pencil Sketch style.`
|
| 328 |
+
* `Convert the video into a Watercolor Painting style.`
|
| 329 |
+
* `Convert the video into a Van Gogh style.`
|
| 330 |
+
* `Convert the video into a Play-Doh style.`
|
| 331 |
+
* `Convert the video into a Claymation style.`
|
| 332 |
+
* `Convert the video into a 3D Chibi style.`
|
| 333 |
+
* `Convert the video into a Ghibli style.`
|
| 334 |
+
* `Convert the video into a Pop Art style.`
|
| 335 |
+
* `Convert the video into an American Cartoon style.`
|
| 336 |
+
* `Convert the video into a Flat Vector Cartoon style.`
|
| 337 |
+
|
| 338 |
+
The training set covers **300+ distinct style names**. Many work; many do
|
| 339 |
+
not. The list above is heavily represented in training. Use the exact
|
| 340 |
+
phrase `Convert the video into a <STYLE> style` β deviations from this
|
| 341 |
+
template degrade quality noticeably.
|
| 342 |
+
|
| 343 |
+
#### What it does *not* do
|
| 344 |
+
|
| 345 |
+
These are honest limits of the training distribution β don't expect them
|
| 346 |
+
to work just because the model is multitask:
|
| 347 |
+
|
| 348 |
+
- **No compositional prompts.** *"Add X and remove Y"*, *"Replace A with B
|
| 349 |
+
and add C"*, etc. are **not** in the training set. Captions combining
|
| 350 |
+
two action verbs are essentially absent (the only ones present are the
|
| 351 |
+
"Remove X and replace with Y" idiom, which is really a single Replace).
|
| 352 |
+
Pure multi-action edits will fall apart β split them into separate runs.
|
| 353 |
+
- **No "change background" as a task.** Background is only used as a
|
| 354 |
+
*positional reference* ("in the background", "on the wall in the
|
| 355 |
+
background"). To swap the entire backdrop, phrase it as a **Replace**
|
| 356 |
+
on a concrete background element, e.g.
|
| 357 |
+
*"Replace the brick wall in the background with a forest at sunset"*.
|
| 358 |
+
Vague prompts like *"Change the background to a beach"* are off-
|
| 359 |
+
distribution and rarely work.
|
| 360 |
+
- **No global colour grade / lighting change.** Only the Style template
|
| 361 |
+
is trained as a global transform. Anything else global (LUT-style
|
| 362 |
+
adjustments, time-of-day swaps without a concrete object) is unreliable.
|
| 363 |
+
|
| 364 |
+
### Which LoRA should I use?
|
| 365 |
+
|
| 366 |
+
| If you want⦠| Use |
|
| 367 |
+
|---|---|
|
| 368 |
+
| Motion transfer (edit first frame externally, model copies motion) | **v0.1 motion transfer** |
|
| 369 |
+
| Multi-task edits (add / remove / replace / style) driven only by prompt | **no-ref multitask r256** (standalone) |
|
| 370 |
+
| Strong identity transfer from a reference image (Add / Replace) | **Ref V2V** |
|
| 371 |
+
|
| 372 |
+
### How to wire the LoopingSampler
|
| 373 |
+
|
| 374 |
+
A single **standard LoRA**, no sidecar, no stacking. Load through one
|
| 375 |
+
ComfyUI LoraLoader before the LoopingSampler. On the sampler:
|
| 376 |
+
|
| 377 |
+
- `editanything_module` β **leave disconnected**.
|
| 378 |
+
- `ref_image` β **leave disconnected**. This LoRA has no reference-image
|
| 379 |
+
path; passing one will just pre-encode tokens that you do not want.
|
| 380 |
+
- `guide_frames` β the guide video.
|
| 381 |
+
- `enable_role_embedding`, `enable_adaln`, `enable_visual_crossattn` β
|
| 382 |
+
all **off** (no module = nothing to inject anyway).
|
| 383 |
+
|
| 384 |
+
---
|
| 385 |
+
|
| 386 |
+
## 3. Reference video-to-video (Ref V2V) β experimental
|
| 387 |
+
|
| 388 |
+
Files (two builds of the same LoRA family β each ships as a `(.standard, .module)` pair):
|
| 389 |
+
|
| 390 |
+
- `edit_anything_reference_v0.1_r128_ref_adaln_proj-role_embedding.standard.safetensors`
|
| 391 |
+
- `edit_anything_reference_v0.1_r128_ref_adaln_proj-role_embedding.module.safetensors`
|
| 392 |
+
- `edit_anything_reference_v0.1_r128_ref_adaln_proj-role_embedding-ref_attn-ref_visual_proj.standard.safetensors`
|
| 393 |
+
- `edit_anything_reference_v0.1_r128_ref_adaln_proj-role_embedding-ref_attn-ref_visual_proj.module.safetensors`
|
| 394 |
+
|
| 395 |
+
### What it is
|
| 396 |
+
|
| 397 |
+
The goal is **add / replace using a reference image** β same vibe as Edit
|
| 398 |
+
Anything v1.0, but with an explicit image as the appearance source instead
|
| 399 |
+
of relying only on the prompt.
|
| 400 |
+
|
| 401 |
+
Trained on **~1600** Add / Replace video pairs. Reference-paired video
|
| 402 |
+
datasets are basically nonexistent, so the dataset had to be built from
|
| 403 |
+
scratch β that is why the sample count is small. **It often fails.** This
|
| 404 |
+
is fully experimental; thousands of training runs went into landing on this
|
| 405 |
+
LoRA layout, and it is still unclear how much it actually helps.
|
| 406 |
+
|
| 407 |
+
### Architecture β why this LoRA has "modules"
|
| 408 |
+
|
| 409 |
+
Trained as a conventional IC-LoRA, plus extra projection branches that try
|
| 410 |
+
to make the reference signal survive across layers:
|
| 411 |
+
|
| 412 |
+
- **`ref_visual_proj`** β projects the reference VAE latent into 32 visual
|
| 413 |
+
memory tokens.
|
| 414 |
+
- **`ref_attn`** β a dedicated cross-attention branch inside each
|
| 415 |
+
transformer block, reading those tokens.
|
| 416 |
+
- **`ref_adaln_proj`** β a global AdaLN bias derived from the reference
|
| 417 |
+
(palette / overall look).
|
| 418 |
+
- **`role_embedding`** β an experimental token bias inspired by some of
|
| 419 |
+
Kijai's tests; whether it actually helps is still unclear.
|
| 420 |
+
|
| 421 |
+
These extra weights are saved alongside the LoRA in a `.module.safetensors`
|
| 422 |
+
sidecar because they are **not standard LoRA adapters** β the regular
|
| 423 |
+
ComfyUI LoRA loader can't consume them, so they need a dedicated node.
|
| 424 |
+
|
| 425 |
+
### How to load
|
| 426 |
+
|
| 427 |
+
| File | What it is | Where it goes |
|
| 428 |
+
|---|---|---|
|
| 429 |
+
| `*.standard.safetensors` | LoRA on `attn1` / `attn2` / `ff` only | Standard ComfyUI LoRA loader |
|
| 430 |
+
| `*.module.safetensors` | `role_embedding`, `ref_adaln_proj`, `ref_visual_proj`, `ref_attn` LoRA adapters | `LTXVEditAnythingModuleLoader` (BFSnodes) |
|
| 431 |
+
|
| 432 |
+
Both files of a pair must be loaded **together** β the LoRA was trained
|
| 433 |
+
against the sidecar adapters and they only make sense as a unit. Do not mix
|
| 434 |
+
`.standard` from one build with `.module` from another.
|
| 435 |
+
|
| 436 |
+
The module file is consumed by the **`π
π
£π
§ LTXV Edit Anything Looping
|
| 437 |
+
Sampler`** node, which was written specifically to:
|
| 438 |
+
|
| 439 |
+
1. Install the `ref_attn` cross-attention branch on every transformer block.
|
| 440 |
+
2. Inject the AdaLN / role / visual cross-attention conditioning at the
|
| 441 |
+
correct points in the model.
|
| 442 |
+
3. Sample long videos in overlapping chunks with the conditioning re-applied
|
| 443 |
+
per chunk.
|
| 444 |
+
|
| 445 |
+
### How to wire the LoopingSampler
|
| 446 |
+
|
| 447 |
+
- Load the `*.standard.safetensors` through a normal ComfyUI LoraLoader
|
| 448 |
+
before the sampler.
|
| 449 |
+
- Load the `*.module.safetensors` through `LTXVEditAnythingModuleLoader`
|
| 450 |
+
and connect its `editanything_module` output to the sampler.
|
| 451 |
+
- On the sampler:
|
| 452 |
+
- `editanything_module` β **the module loader output** (required).
|
| 453 |
+
- `ref_image` β **the reference image** (required β this is what
|
| 454 |
+
`Add` / `Replace` will insert).
|
| 455 |
+
- `guide_frames` β the source video to edit.
|
| 456 |
+
- `enable_adaln` β **on** (defaults match training).
|
| 457 |
+
- `enable_visual_crossattn` β **on** for the 4-extras build; off (or
|
| 458 |
+
will be a no-op) for the 2-extras build.
|
| 459 |
+
- `enable_role_embedding` β **off** for the 4-extras build (training
|
| 460 |
+
config disabled it). On if you're loading the 2-extras build alone.
|
| 461 |
+
|
| 462 |
+
Missing `ref_image` here silently disables AdaLN and the visual
|
| 463 |
+
cross-attention β the sampler will warn in the log.
|
| 464 |
+
|
| 465 |
+
### Which build to use
|
| 466 |
+
|
| 467 |
+
- **`ref_adaln_proj-role_embedding`** β the original training. Only ships
|
| 468 |
+
the two side-channel modules.
|
| 469 |
+
- **`ref_adaln_proj-role_embedding-ref_attn-ref_visual_proj`** β the
|
| 470 |
+
continuation. Adds the visual cross-attention branch and its projector on
|
| 471 |
+
top.
|
| 472 |
+
|
| 473 |
+
It is genuinely **not clear yet** whether the extra branches help over the
|
| 474 |
+
plain LoRA. Both builds are honest experiments. Try both, decide for your
|
| 475 |
+
own use case, and please share findings.
|
| 476 |
+
|
| 477 |
+
### Reading the layers
|
| 478 |
+
|
| 479 |
+
For anyone who wants to understand what each layer in the Ref V2V
|
| 480 |
+
checkpoint does:
|
| 481 |
+
|
| 482 |
+
- [`lora_layers_reference.md`](./lora_layers_reference.md) β full tensor
|
| 483 |
+
inventory of both builds.
|
| 484 |
+
- [`lora_layers_impact.md`](./lora_layers_impact.md) β what each branch
|
| 485 |
+
contributes at inference and which inference knob (`adaln_scale`,
|
| 486 |
+
`ref_context_scale`, `ref_token_scale`, `ref_start_block`,
|
| 487 |
+
`ref_end_block`, etc.) maps back to which training default.
|
| 488 |
+
|
| 489 |
+
---
|
| 490 |
+
|
| 491 |
+
## Prompt examples
|
| 492 |
+
|
| 493 |
+
The two LoRAs were trained on very different caption styles. Match the
|
| 494 |
+
style of whichever LoRA you're using β straying outside the training
|
| 495 |
+
distribution is the fastest way to get garbage out.
|
| 496 |
+
|
| 497 |
+
### Edit Anything v0.1 β standard editing
|
| 498 |
+
|
| 499 |
+
The stage-1 dataset uses short imperative captions describing one or two
|
| 500 |
+
edits. Use the same shape at inference. Examples drawn from the training
|
| 501 |
+
distribution:
|
| 502 |
+
|
| 503 |
+
- *"Replace the stone statue of a man on the left with a young woman in a
|
| 504 |
+
green dress."*
|
| 505 |
+
- *"Add a black labrador retriever sitting beside the woman on the bench."*
|
| 506 |
+
- *"Remove the teacher from the classroom."*
|
| 507 |
+
- *"Alter the cap's colour from modern black to deep maroon."*
|
| 508 |
+
- *"Replace the fresh citrus-green background with a wooden desk."*
|
| 509 |
+
- *"Add faint tire tracks across the snow behind the car."*
|
| 510 |
+
- *"Add a black statue, a blue camera, a cyan towel, a red guitar and a
|
| 511 |
+
pink backpack to the lakeside pier."*
|
| 512 |
+
|
| 513 |
+
Tips:
|
| 514 |
+
|
| 515 |
+
- Imperative verbs: **Add / Replace / Remove / Alter / Change**.
|
| 516 |
+
- When replacing, **describe both** the original and the new subject so the
|
| 517 |
+
model can localise the edit.
|
| 518 |
+
- Keep captions short and concrete. Long flowery prose hurts.
|
| 519 |
+
|
| 520 |
+
### Edit Anything v0.1 β motion transfer
|
| 521 |
+
|
| 522 |
+
Workflow:
|
| 523 |
+
|
| 524 |
+
1. Pick a guide video.
|
| 525 |
+
2. Edit **only the first frame** externally (Flux Klein or any
|
| 526 |
+
capable image-edit model) to introduce the new subject in the desired
|
| 527 |
+
pose and position.
|
| 528 |
+
3. Feed the edited frame as the first frame of the input and the original
|
| 529 |
+
guide as motion source.
|
| 530 |
+
4. The prompt should describe **the inserted subject and the action being
|
| 531 |
+
preserved**.
|
| 532 |
+
|
| 533 |
+
Examples:
|
| 534 |
+
|
| 535 |
+
- *"Replace the standing man holding the umbrella with a woman in a red
|
| 536 |
+
coat holding the same umbrella, walking across the puddles."*
|
| 537 |
+
- *"Add a tabby cat curled up in the armchair while the man in the
|
| 538 |
+
background keeps reading."*
|
| 539 |
+
- *"Replace the runner in the blue jersey with a man wearing a white shirt
|
| 540 |
+
and grey shorts running along the same path."*
|
| 541 |
+
|
| 542 |
+
Limits: fast or chaotic motion will fail; the inserted subject should
|
| 543 |
+
occupy roughly the same region/scale as what it replaces.
|
| 544 |
+
|
| 545 |
+
### Reference V2V (Ref V2V) β Add and Replace
|
| 546 |
+
|
| 547 |
+
These captions are real samples from the ~1600-pair training set. They
|
| 548 |
+
describe the **target scene after the edit** in detail. The reference
|
| 549 |
+
image carries the *appearance* of the inserted subject; the caption
|
| 550 |
+
carries *position, pose, action, and surrounding context*.
|
| 551 |
+
|
| 552 |
+
**Add task** (the reference image holds the new subject):
|
| 553 |
+
|
| 554 |
+
- *"Add a middle-aged man with curly grey hair, a beard and glasses,
|
| 555 |
+
wearing a blue quarter-zip sweater, on the right side of the frame,
|
| 556 |
+
standing in front of a raw cut of meat on a tray."*
|
| 557 |
+
- *"Add a light-coloured small boat with dark seats and an outboard motor
|
| 558 |
+
floating in the water."*
|
| 559 |
+
- *"Add an open book filled with colourful pencils in the woman's hands."*
|
| 560 |
+
- *"Add a silver metallic bucket on the table in front of the blonde
|
| 561 |
+
character, with her hands stirring a mixture inside."*
|
| 562 |
+
- *"Add two miniature dolls, one blonde and one brunette, dressed in
|
| 563 |
+
patterned clothing, sitting at a small table with teacups and small
|
| 564 |
+
white vases on the countertop."*
|
| 565 |
+
|
| 566 |
+
**Replace task** (the reference image holds the new subject; the caption
|
| 567 |
+
also describes what is being replaced):
|
| 568 |
+
|
| 569 |
+
- *"Replace the standing kangaroo holding the bicycle handlebars with a
|
| 570 |
+
man wearing a white t-shirt, light brown shorts and a yellow cap,
|
| 571 |
+
holding the bicycle handlebars."*
|
| 572 |
+
- *"Replace the stone statue of a man on the left side with a young woman
|
| 573 |
+
in a green dress."*
|
| 574 |
+
- *"Replace the wooden barrel near the entrance with a large brown leather
|
| 575 |
+
suitcase."*
|
| 576 |
+
|
| 577 |
+
Tips for Ref V2V:
|
| 578 |
+
|
| 579 |
+
- **Describe the inserted subject in full**, even though the reference
|
| 580 |
+
image is the source of truth β the text path drives placement and pose.
|
| 581 |
+
- For *Replace*, **also describe what is being replaced** so the model can
|
| 582 |
+
match the spatial region.
|
| 583 |
+
- Keep the inserted subject roughly in the same scale and region as what
|
| 584 |
+
it replaces.
|
| 585 |
+
- The captions in the training set average ~25β40 words β aim for that
|
| 586 |
+
range. Single-sentence captions like *"Add a man"* are far too sparse
|
| 587 |
+
and will fail.
|
| 588 |
+
|
| 589 |
+
---
|
| 590 |
+
|
| 591 |
+
## Inference tips (applies to all models)
|
| 592 |
+
|
| 593 |
+
**CFG matters a lot here.** The default workflow runs with the LTX-2.3
|
| 594 |
+
**distilled / acceleration LoRAs** for fast 4β8 step sampling, which
|
| 595 |
+
locks **CFG = 1.0**. That's fine for casual runs, but at CFG 1 the model
|
| 596 |
+
follows the prompt loosely β you get the reference image to "show up"
|
| 597 |
+
but the edit instruction itself is only weakly enforced.
|
| 598 |
+
|
| 599 |
+
**For harder prompts, raise CFG above 1.0.** This means dropping (or
|
| 600 |
+
weakening) the distilled / acceleration LoRAs and going back to a normal
|
| 601 |
+
sampler with more steps β significantly slower, but the model follows
|
| 602 |
+
the prompt much more closely. Trade-off:
|
| 603 |
+
|
| 604 |
+
Other knobs:
|
| 605 |
+
|
| 606 |
+
- If the model is **ignoring the prompt** (edit isn't being applied, the
|
| 607 |
+
reference is barely showing up, the style transfer is faint), raising
|
| 608 |
+
CFG is the single most common fix. Go up to 6β8 if needed.
|
| 609 |
+
- If results look **over-saturated, plasticky, or motion is freezing**,
|
| 610 |
+
CFG is too high β pull back toward 3β4 or re-enable the distilled LoRA
|
| 611 |
+
for CFG 1 if you don't actually need stronger prompt adherence.
|
| 612 |
+
- Ref V2V in particular benefits from being more aggressive with CFG when
|
| 613 |
+
the reference identity isn't transferring cleanly.
|
| 614 |
+
- Combine CFG tuning with the LoRA-specific knobs from each section
|
| 615 |
+
(`adaln_scale`, `ref_context_scale`, `ref_token_scale` for Ref V2V;
|
| 616 |
+
prompt rewriting for v0.1 / no-ref).
|
| 617 |
+
|
| 618 |
+
Treat CFG as a real knob, not a constant β and be ready to give up some
|
| 619 |
+
speed when you actually need the edit to land.
|
| 620 |
+
|
| 621 |
+
---
|
| 622 |
+
|
| 623 |
+
## ComfyUI nodes
|
| 624 |
+
|
| 625 |
+
All recommended inference paths run through the **BFSnodes** custom node
|
| 626 |
+
set. For now BFSnodes is the only place these nodes live; once they
|
| 627 |
+
stabilise they may move elsewhere.
|
| 628 |
+
|
| 629 |
+
Specific nodes used by these LoRAs:
|
| 630 |
+
|
| 631 |
+
- `LTXVEditAnythingApply` β load the LoRA + extras and patch the model.
|
| 632 |
+
- `π
π
£π
§ LTXV Edit Anything Looping Sampler` β sampler that injects role /
|
| 633 |
+
AdaLN / visual cross-attention and handles long videos in chunks.
|
| 634 |
+
- `LTXVEditAnythingModuleLoader` β load the `*.module.safetensors` sidecar.
|
| 635 |
+
|
| 636 |
+
---
|
| 637 |
+
|
| 638 |
+
## Status
|
| 639 |
+
|
| 640 |
+
Released as experimental research artefacts. Expect failures, do not
|
| 641 |
+
deploy, and please report what works and what doesn't.
|
| 642 |
+
|
| 643 |
+
---
|
| 644 |
+
|
| 645 |
+
## Credits
|
| 646 |
+
|
| 647 |
+
If you use these models β in a project, a demo, a paper, a video, a tweet,
|
| 648 |
+
a workflow, anything β **please credit my work**. These checkpoints are the
|
| 649 |
+
result of weeks of research, dataset building, and training runs, and that
|
| 650 |
+
effort is what makes any of it usable. Crediting the source is the bare
|
| 651 |
+
minimum that keeps open research like this sustainable.
|
| 652 |
+
|
| 653 |
+
**Author:** Alisson Pereira dos Anjos ([@Alissonerdx](https://huggingface.co/Alissonerdx))
|
| 654 |
+
|
| 655 |
+
Suggested attribution:
|
| 656 |
+
|
| 657 |
+
> Edit Anything LoRAs by Alisson Pereira dos Anjos
|
| 658 |
+
> ([huggingface.co/Alissonerdx/EditAnything](https://huggingface.co/Alissonerdx/EditAnything)).
|
| 659 |
+
|
| 660 |
+
Links back to this repository are appreciated wherever you publish results.
|