Image Oasis

Oasis Suite

One ComfyUI pack with four nodes that share the same player language, save habits, and UI patterns. Install once; everything lands under custom_nodes/ComfyUI-Image-Oasis/.

| Node | Class id | Folder | Docs | |------|----------|--------|------| | Image Oasis | ImageOasis | image_oasis/ | image_oasis/image_oasis_README.md | | Video Oasis Viewer | VideoOasisPreview | video_oasis/ | video_oasis/video_oasis_README.md | | LTX Oasis | LTX23Oasis | ltx23_oasis/ | ltx23_oasis/ltx23_oasis_README.md | | Audio Oasis | AudioOasis | audio_oasis/ | audio_oasis/audio_oasis_README.md |

Frontends live in web/ (image_oasis.js, video_oasis.js, ltx23_oasis.js, audio_oasis.js). License: GPL-3.0-or-later for the whole pack (see LICENSE).

---

What's new in 1.7

LTX 2.5 (LTX Oasis)

The architecture picker gains LTX 2.5 22B (Distilled) alongside 2.3. It is one registry entry rather than a new node, so the class id stays LTX23Oasis and existing workflows are untouched - the dropdown decides which model family a run targets.

The one thing to know when switching: 2.5 takes a single text encoder. It ships Gemma-4-12B with the LTX projection baked into one file, where 2.3 needs Gemma-3 plus a separate projection file, so the Model section drops to one text-encoder picker. Its defaults also differ - 24 fps and 960x544 against 2.3's 25 fps and 1280x720 - while the sigma schedule, sampler and CFG are the same. Everything else (Prompt Beats, guides, audio, motion context, the player and scene bar) behaves identically on both.

2.5 has no spatial upscaler of its own yet, so Spatial Upsample on a 2.5 run loads the 2.3 upscaler, which is what the official 2.5 two-stage workflow does.

The LTX node is now "LTX Oasis"

With 2.5 in the picker the node is no longer 2.3-only, so its header reads LTX Oasis. Display name only: the class id stays LTX23Oasis, the folder stays ltx23_oasis/, routes stay /ltx23_oasis/*, and presets and themes stay under user/ltx23_oasis/. Saved workflows keep loading, and a node you renamed by hand keeps your title.

Create Movie audio (Video Oasis Viewer and LTX Oasis)

Joining clips used to drop a short hole into continuous ambience at every cut, and audio walked progressively late against picture down a long scene bar. Both came from measuring audio the wrong way: every clip's AAC encode pads its tail out to a whole block, and clip files run fractionally longer than their own video. Audio is now cut to exactly the length each clip's frame count calls for, so the two are measured off the same ruler.

The remaining artifact was the incoming clip's first few frames arriving weak, which a trim cannot recover because it is real audio, just quiet. The samples either side of a join are now rebuilt by crossfading the two clips' own interiors played backwards, which cannot click and does not change the sample count. Create Movie reports what it did to each clip in the ComfyUI console.

---

What's new in 1.6

Motion context for chained clips (LTX2.3 Oasis)

Continue from viewed video could only ever see one frame of the previous clip. A still carries pose but not motion - a leg mid-stride looks the same swinging forward or back, a camera frozen mid-pan gives no hint which way it was moving - so every join was a fresh guess and chained clips lurched.

1.6 pins a window of the previous clip in front of the timeline as frozen frames (9 / 17 / 25 / 33 / 41 / 49; 25 is one second at 25fps, 49 is just under two). The model reads direction, speed, gait phase and camera drift out of them. The window is cropped after decode, delivered frame counts are unchanged, and audio is untouched - File mode pads the window with silence so an uploaded track is muxed exactly as supplied, Generate mode carries the previous clip's tail audio so ambient beds and music continue across the cut.

The tail source is now read from the file in the viewer rather than an in-memory tensor, so clips loaded from output/ chain exactly like fresh renders.

Audio Oasis (new node)

A fourth node joins the suite. Load an mp3/wav/flac/m4a, chop it into segments on a waveform, save them as numbered files, and feed the result to LTX2.3 Oasis's audio-driven-video slot or any other AUDIO-input node.

• Waveform, duration, sample rate, channels and peak/RMS decode in the browser on upload, no server round trip
• Click to drop a chop point, drag to slide, double-click or right-click to remove; split every N s bulk-places evenly spaced points you can still edit afterward
• Grid & Snap keeps cuts legal for LTX: snap modes 8n+1 (default), frames, or off, with Re-snap after an FPS change. Every segment row shows its frame count and flags any length LTX would reject
• Save segments writes numbered files under input/audio_oasis/<track>/ in the source format, re-encoded so cuts land sample-accurate rather than on the nearest frame boundary, plus a manifest.json
• Load saved set restores track, points, FPS and segments from that manifest, so a chop survives restarts even if the workflow was never saved
• Drag a segment's grip straight onto LTX2.3 Oasis's audio slot, or pick a segment (or Use full track as output) to set what the AUDIO socket carries
• Analyze gives BPM and estimated key via optional librosa
• Its own theme, independent of Image Oasis and LTX2.3 Oasis

Full detail: audio_oasis/audio_oasis_README.md (and the in-node Help pane, fed by audio_oasis/audio_oasis_help_content.md).

Video Oasis Viewer

The scene bar holds 48 clips, up from 24. LTX2.3 Oasis's bar shares the constant and gets the same capacity.

---

What's new in 1.5

v1.5 turns Image Oasis into a small suite. Two new nodes ship in this pack for the first time (they were never a public release on their own):

Video Oasis Viewer (new)

A preview-first Save Video replacement. Incoming VIDEO encodes to temp and plays in-node; nothing hits your output folder until you press Save.

• Scrub / frame-step / mute / speed / lightbox (scroll zoom, drag pan)
• Frame drag onto other nodes' image inputs (Load Image, refs, LTX guides, ...)
• Playback: off → loop → cycle (cycle walks the scene bar like a dailies reel)
• Scene bar (up to 24, raised to 48 in 1.6): click to recall, delete, long-press reorder, + load from output/, Save (lossless copy + workflow metadata)
• Clip: mark in [ / out ] then Clip - trimmed file lands in the bar
• Create Movie: concat every saved bar clip, result lands back in the bar (stream-copy when possible, re-encode when Clip or mismatched params require it)
• Encode / Save section (LTXO-matching): format / codec / quality / save prefix
• Theme follows LTX2.3 Oasis (not Image Oasis)

Full detail: video_oasis/video_oasis_README.md.

LTX2.3 Oasis (new)

All-in-one LTX 2.3 video generation in the Image Oasis UI shape - model pick, prompt enhancer, LoRA stack, Start Frame / Prompt Beats, audio modes, sigmas, spatial upsample - with the same player, scene bar, Clip, Create Movie, and encode/save path as Video Oasis Viewer (uses the in-pack encode path; no extra video pack required).

• Text→Video and Image→Video, plus Continue from viewed video chaining
• Prompt Beats (local text + optional guide images on a timeline)
• Audio: Off / Generate / File (audio-driven video)
• Scene bar persistence across ComfyUI tab switches and reloads

Full detail: ltx23_oasis/ltx23_oasis_README.md (and the in-node Help pane, fed by ltx23_oasis/ltx23_oasis_help_content.md).

Image Oasis (carried forward)

Same all-in-one image node as 1.4.x, plus the recent history-strip / bypass / CivitAI-hash work. Architecture registry, enhancer, LoRAs, refiner, upscale - unchanged class id ImageOasis. See image_oasis/image_oasis_README.md for the full feature list and older "what's new" notes (v1.1-1.4).

Pack-level changes

• Layout: image_oasis/, video_oasis/, ltx23_oasis/ + shared web/
• Whole pack is GPL-3.0-or-later (LTX Director code vendored under ltx23_oasis/vendor/ requires it)
• Stable class ids so workflows keep loading: ImageOasis, VideoOasisPreview, LTX23Oasis

---

Install

1. Place (or update) this folder at ComfyUI/custom_nodes/ComfyUI-Image-Oasis/.
2. Install Python deps from the pack root: pip install -r requirements.txt
3. Restart ComfyUI, then hard-refresh the browser (Ctrl+F5) so the frontend scripts reload.

Optional (Image Oasis / LTX enhancer): llama-cpp-python with a CUDA/Metal build for the GGUF prompt enhancer - see image_oasis/image_oasis_README.md.

Optional (LTX): ComfyUI-GGUF for GGUF diffusion; ComfyUI-KJNodes for LTX2 NAG when CFG is 1 and a negative prompt is set.

Optional (Audio Oasis): librosa for BPM/key Analyze - listed commented-out in requirements.txt because it pulls in numba/llvmlite. Chop / save / drag all work without it.

Presets / themes:

• Image Oasis → ComfyUI/user/image_oasis/
• LTX Oasis → ComfyUI/user/ltx23_oasis/
• Audio Oasis → ComfyUI/user/audio_oasis/

---

Credits

• Execution timer pattern adapted from crt-nodes.
• Krea 2 conditioning rebalance: nova452 / huwhitememes (Apache-2.0).
• LTX Oasis vendors PromptRelay / patches from WhatDreamsCost-ComfyUI (LTX Director), GPL-3.0-or-later - see ltx23_oasis/vendor/.
• "Accessibility tool for the nodally challenged" - PheebyKatz.

License

GPL-3.0-or-later. See LICENSE.