v1.5.0 – keyword output and filter_mode OFF for the Filtered random loader

Added

• keyword output (Anima Filtered Random LoRA Loader): the keywords the selected LoRA actually matched, joined with _
  • AND: aaa bbb → aaa_bbb
  • OR: aaa bbb → aaa or bbb
  • Quoted phrases: "aaa bbb" ccc (AND) → aaa_bbb_ccc
  • Empty when filter_mode is OFF, keyword_filter is empty, or no LoRA matched
• filter_mode OFF: turns the keyword filter off without deleting the keywords you typed

Compatibility

• keyword is added as the last output, so existing workflow connections keep working
• LoRA selection and all other outputs are unchanged; the same seed gives the same result as v1.4.0

v1.4.0 - Random LoRA loaders: clean positive_text, new lora_text output

Changed

positive_text no longer contains <lora:...> tags (Anima Random LoRA Loader / Anima Filtered Random LoRA Loader)

positive_text is now the additional prompt plus trigger words only — exactly what gets encoded into the positive CONDITIONING output. It can be fed straight into a text encoder without the <lora:...> tags being read as text.

The previous content, tags included, now comes from a new lora_text output, added as the last slot.

⚠️ What to check in existing workflows

• If positive_text fed a text encoder: nothing to do. Since the tags are no longer read as noise, results with the same seed may differ.
• If positive_text was used to record the LoRA syntax (saved prompts, metadata): reconnect that link to lora_text.

All other outputs keep their positions and types, so no other connections are affected.

Fixed

• Filtered Random LoRA Loader: when two selected LoRAs without trigger words were adjacent, a stray , , was left in the encoded prompt after removing the tags. Runs of commas are now collapsed.

Upgrading

Restart ComfyUI. See the Node 5 & 6 output description in the README for details.

v1.3.0 - ControlNet support on 40/52-block models (LLLite & VACE)

New

Apply Anima ControlNet-LLLite (Auto Remap)

A version of kohya-ss's "Apply Anima ControlNet-LLLite (sd-scripts)" that works on 28-, 40- and 52-block models. LLLite weights trained on Anima-Base previously stopped with a depth_embed slices missing error on larger models; this node renumbers the blocks in the weights to match the connected model before loading.

• Behaves exactly like upstream on a 28-block model
• Optional extend_to_new_layers applies control on the newly-inserted blocks too
• kohya-ss's package is not required (the needed code is bundled, Apache 2.0)

Anima VACE ControlNet Remap

VACE-type ControlNets (TaihoC's Depth, khanghy1000's Canny) run on 40/52-block models without any error but inject at the wrong blocks, because the injection points are fixed to the 28-block layout. This node sits between the loader and the apply node and moves them to the right blocks (e.g. 0/7/14/21 → 0/10/20/31 on 40 blocks).

Requires the fix/anima-vace-hardening branch of PineCookie's ComfyUI-Advanced-ControlNet fork. Anima-Remap itself still loads normally without it.

Fixed

• fp8 weights: LoRAs and LLLite files saved in fp8 failed with "mul_cpu_reduced_float" not implemented for 'Float8_e5m2' when extend_to_new_layers was ON. The extension is now computed in float32.

Changed

• Random LoRA loaders aligned with RandomLoRALoader v1.3.0: the candidate list is sorted before selection (so a seed reproduces across machines), and .pt / .ckpt files no longer log a spurious metadata error.

⚠️ Because of the sorting, an existing workflow may pick a different set of LoRAs after upgrading, even with a fixed seed.

Upgrading

Restart ComfyUI. See the README sections for Node 7 and Node 8 for details and example wiring.

v1.2.0 - LoRA name autocomplete with previews

New

LoRA name autocomplete in the Tag Loader text box

While typing directly into the text box of the Anima LoRA Tag Loader (regular and Extended), type 3 or more characters of a LoRA filename to get a list of partial matches.

• Highlight a candidate with the arrow keys or mouse to see its preview and trigger words
• Press Enter / Tab or click to insert <lora:filename:1>, trigger words in place of what you typed
• Previews support images and video (mp4 / webm, played muted on a loop), using LoRA Manager's .preview.* files or plain <name>.png/.jpg/...
• Trigger words come from .metadata.json (LoRA Manager), .civitai.info / .info (Civitai Helper), or the embedded modelspec.trigger_word
• The bare filename is inserted; a subfolder is added only when the same filename exists in more than one folder

Connecting text from an upstream node works exactly as before — the autocomplete only appears when typing into the node's own text box.

See the "LoRA name autocomplete" section of the README for details.

Upgrading

Restart ComfyUI, then hard-reload the browser (Ctrl+F5) so the new JavaScript is loaded.

No new dependencies and no changes to existing node behaviour or workflows.

v1.1.0-Add settings hash to remap cache filenames, add remap_info output, fix LoRA name resolution

Add settings hash to remap cache filenames, add remap_info output, fix LoRA name resolution

• Cache filenames now embed a hash of manifest/extend settings, so
  changing settings no longer silently reuses a stale cache
• New remap_info STRING output on both LoRA nodes
• Fix LoRA tag names with an extension and/or subfolder failing to
  resolve (issue #1); accept Windows-style separators and any casing
• Warn with diagnostic context when a name can't be matched
• Random loaders now follow symlinks when scanning folders, matching
  ComfyUI's own behaviour, with loop detection
• Document the above in both READMEs