1.1.3 Hangeul — prompt autocomplete

1.1.3 Hangeul — Prompt autocomplete

RS Workspace reuses the original ComfyUI prompt textarea so an installed,
enabled Custom Scripts AutoCompleter or Autocomplete-Plus can keep providing
its suggestions. Autocomplete++ is also detected; its full UI integration has
not been exercised. rgthree Power Prompt is not a global autocomplete provider.

When no known provider is enabled, RS supplies lightweight offline suggestions
from bundled tags, prompt tags already in the graph and local wildcard filenames.
Use Up/Down to select, Tab/Enter to insert, Escape to dismiss, and Ctrl+Space to
open suggestions. Weighted text and trailing prompts are retained. IME composition
is protected, and literal typography inputs are excluded.

An existing user dictionary can be copied to web/autocomplete.txt inside this
extension. Custom Scripts tag,count text and A1111 CSV files are supported.
Counts prioritize matches, aliases insert canonical tags, and parentheses are
escaped. The optional dictionary is personal data: it is ignored by Git and
is not included in the release. No third-party extension or dictionary is
automatically downloaded.

Update the extension, restart ComfyUI when idle, then hard-refresh the browser
with Ctrl+Shift+R (macOS: Cmd+Shift+R). Built-in suggestions can be switched off
with RS: Built-in prompt autocomplete in ComfyUI settings. External providers
retain their own enable settings.

Validation scope

Token insertion, provider dispatch, dictionary parsing, line endings, wildcard
names, weighted prompts and literal text exclusion have regression coverage.
Actual browser checks of the fallback, original textarea reuse, tab changes,
popup cleanup, and Custom Scripts/Autocomplete-Plus integration are recorded in
prompt autocomplete validation. External
provider data used in integration checks was represented by small fixtures;
compatibility with every third-party configuration is not claimed. This UI
release does not change model inference or claim new image-generation tests.

Release checks: Python unittest discovery ran 427 tests successfully (413 passed,
14 skipped) against the local ComfyUI runtime. All 20 JavaScript regression
scripts passed, including camera Python parity checks.

1.1.2 Hangeul — workspace snapshot fix

1.1.2 Hangeul — Reactive workflow snapshot fix

RS Workspace now snapshots workflow and widget settings through the JSON format
used by workflow files and the /prompt API. Native structuredClone rejects
otherwise serializable reactive JavaScript Proxy objects, which can prevent
app generation or model-setting persistence with a DataCloneError.

Resolved wildcard prompts are detached before source templates are restored.
Queued workflow provenance and saved settings do not share mutable nested
objects with the live graph. Serialization failures still restore the templates
and are reported rather than submitting a partial request. App errors also log
the original stack in the browser console to help diagnose other reports.

Update

Update to 1.1.2, then hard-refresh the ComfyUI browser tab with
Ctrl+Shift+R (macOS: Cmd+Shift+R). A server already serving the new files
can serve this JavaScript fix without interrupting its current generation queue.
Use the normal restart procedure after a full extension update when idle.

Evidence and limits

The old snapshot path reproduced DataCloneError with nested Proxy objects.
The revised path passed browser and Node.js regressions covering a reactive
workflow, reactive wildcard metadata, model-setting snapshots, template recovery
on success and failure, independent nested copies, Unicode and large safe
integer seeds. All JavaScript regression scripts passed locally.

The original report contained only the exception message, without a stack,
frontend version or action that triggered it. These tests prove this app-side
failure mode and its fix; they do not prove the reporter used this path. The
exception is a browser serialization failure and does not identify Windows or
any other operating system. This patch does not alter model inference or claim
new image-quality validation.

1.1.1 Hangeul — 3D pose guidance in the building pipe

1.1.1 Hangeul — 3D pose guidance in the building pipe

Configure learned pose guidance directly in RS 3D Pose Masks. Connect its
building_pipe to the character builder and Finalize, then use the normal RS
sampler. No MODEL, VAE or patch-loader connections are needed on the 3D node.

The sampler uses the model and VAE it already receives to route compatible
Anima 2B, Krea 2 and MOSS guidance. Required pose weights are downloaded when
needed. Unsupported families and Anima 2.9B retain masks and report the reason.

Update and use

1. Update the extension to 1.1.1 and restart ComfyUI.
2. Hard-refresh the browser with Ctrl+Shift+R (macOS: Cmd+Shift+R).
3. In RS 3D Pose Masks, edit the scene, choose auto, and set strength and
   start/end. The building pipe can carry prompts and positions in either order.
4. Connect to Finalize and an RS sampler. The RS Workspace app now uses this
   same route automatically.

Saved older 3D nodes retain masks_only until explicitly changed. Existing app
graphs migrate their old generated helper settings into the 3D editor. The old
MODEL-input pose adapter remains deprecated for older saved workflows; remove it
when enabling the new route so the model is not patched twice.

Validation and limitations

The Python regression suite ran 427 tests: 413 passed and 14 skipped. The final
targeted pose recheck ran 51 tests: 49 passed and 2 skipped. All 18 JavaScript
regression scripts passed. An actual Anima 2B image was generated on RTX 5060 Ti
at 768×1024, 30 steps, Euler ancestral / SGM uniform through the building pipe.
The PNG was opened in ComfyUI to verify its restored controls and connections.

Detailed trial and scope ·
3D setup guide ·
Loadable workflow PNG.

Krea 2/MOSS learned-guidance routing is covered by regressions; this patch does
not claim new image-quality trials for them. Their learned guidance requires
full resolution. Learned guidance cannot currently be combined with image
references. Pose guidance is approximate and does not guarantee exact joints,
hands, identity or anatomy.

1.1.0 Hangeul — RS Workspace

RS Workspace adds an experimental app screen inside ComfyUI. Select an installed model and use character prompts, LoRAs, masks, 3D poses, camera, references and face detailing through the existing RS nodes. Model-specific loaders, starting settings and supported feature routes are selected automatically.

How to open

1. Update RS MultiCharacter Sampler to 1.1.0 through ComfyUI-Manager, or run git pull --ff-only in an existing Git installation.
2. Restart the ComfyUI server, then hard-refresh the browser (Ctrl+Shift+R, macOS Cmd+Shift+R).
3. Click RS Workspace at the bottom right. The alternative menu command is Edit → RS Model Workspace (Experimental) where available.
4. Select an installed model, edit the scene and click 생성 (Generate). 노드 그래프 (Node graph) returns to the editable workflow.

No separate app installation is required. Korean labels are currently used in the app; the English guide maps the main controls. Models, encoders and VAEs are installed separately. Missing files and unsupported feature combinations are reported before generation.

Guides and examples

• English app guide and troubleshooting
• 한국어 전체 가이드 — 설치부터 모든 탭·와일드카드·저장·복원까지
• Validation scope and limitations
• Download RS-Workspace-Wildcard-Example.zip below for the workflow PNG, background TXT and instructions. The standalone PNG is also attached.

Included

• Model-aware routing and per-model settings restoration.
• Existing mask/position/3D/camera editors, character and global LoRAs, references, interaction controls, progressive sampling and face detailing.
• Wildcard TXT search/insertion, seeded preview and expansion; PNGs keep resolved prompts and original templates.
• Separate workflow tabs, saved workflow recovery, generation progress and results.
• Preserve a customized output filename prefix when reopening the app.

Checks and limits

Python: 421 tests, 414 passed, 7 skipped. JavaScript: 19 test files passed. Existing real app-route executions cover Anima 2B/2.9B, SDXL, SD1.5, Krea 2, MOSS and tested FLUX variants; actual browser generation and PNG restoration were verified separately. These are not an exhaustive visual quality matrix.

The app remains experimental. One click queues one image request. GGUF inventory and arbitrary encoder names are not automatically handled. Existing model-specific pose/reference/progressive restrictions and unstable SD1.5 quality remain documented. Use the node graph for unsupported app setups.

1.0.11 Hangeul — Image to Editable 3D Pose

1.0.11 Hangeul

2026-10-30 (Asia/Seoul).

• Image → editable 3D: drop/select a single-person PNG, JPEG or WebP on
  the 3D pose node. It queues SAM 3D Body estimation locally, saves the estimated
  scene and opens the editor automatically. No extra image/estimator wiring is
  needed. The existing node and output socket order are preserved.
• Source framing: retain image aspect ratio, camera and crop. Only the body
  inside the source frame enters silhouettes, pose maps and editable geometry;
  off-frame inferred legs are excluded. Hidden joints/depth within the frame
  remain estimates, so use this as a composition guide.
• Import while editing: a one-person scene adopts source framing. In a group,
  replace only the selected person, preserving the other people and scene camera.
  Import applies automatically; later mouse edits use Apply to node. Stale
  inference results cannot overwrite a scene changed while estimation was running.
• Local pose library: choose a catalogue pose, randomize the selected person,
  or randomize each person independently. Existing placement and prompt IDs stay
  fixed. Samples remain local; no downloaded pose data is included in GitHub or
  Registry packages. Randomization happens when clicked, not during sampling.
• Requirements and notices: image estimation needs a ComfyUI core providing
  SAM3DBody_Loader / SAM3DBody_Predict. First use downloads the pinned 2.83 GB
  model. The separate SAM License is included with the third-party notices.
  Manual 3D posing and masks do not need this inference model.

Usage

Restart ComfyUI after updating and hard-refresh its browser tab. Add RS 3D Pose
Masks (Experimental), select/drop an image on Image → 3D, then click a joint
and drag its rotation rings. Move person changes placement; dragging empty
space orbits the camera. Finish with Apply to node. Existing pose-adapter
MODEL and building_pipe connections still carry different signals and both are
needed for trained pose guidance. See the full guide.

한국어: 이미지 한 장을 넣으면 자동 추정 후 마우스로 수정할 수 있는 3D 편집기를
엽니다. 반신은 보이는 화면 범위를 유지하고 화면 밖 하체를 출력하지 않습니다.
여러 명을 편집할 때는 선택 인물만 교체합니다. 로컬 포즈 목록의 랜덤 선택도
추가했습니다. 샘플 파일은 로컬에만 보관하며 패키지에는 포함하지 않습니다.

Verification scope

Release checks and installed-runtime results are recorded in
1.0.11 validation. These import/editor checks
do not establish exact reconstruction, reliable hands/contact, or new generation
quality. The earlier Krea/MOSS audit remains
the separate image-generation evidence.

Release verification

• Main commit: 7d268c0a3abfc9f81d10582afda007a226684fc5.
• Linux / Windows compatibility CI: all four jobs passed.
• Comfy Registry publishing workflow: passed; registry service validation is separate.

1.0.10 Hangeul — Registry package correction

1.0.10 Hangeul

2026-10-30. Corrected Registry package following 1.0.9.

• Includes the editable character STRING sources and connections in all five
  Krea/MOSS workflow PNGs. Frontend workflows now preserve the same captions
  as their embedded API requests.
• Includes the explicit update-ComfyUI error when an older core cannot provide
  scheduled MOSS pose conditioning. Disabled pose and Krea passthrough remain
  available on their supported paths.
• Aligns the package manifest, runtime __version__, five README release links
  and current guide at 1.0.10. A regression checks version agreement.
• Publishes these corrections under a new Registry version because the already
  submitted 1.0.9 package cannot be overwritten.

Pose routing and image-generation behavior are unchanged from the corrected
GitHub 1.0.9 release. See 1.0.9 feature notes, the
36-image audit and the
Korean guide for setup and known limitations.

한국어: Registry와 GitHub의 배포 내용을 맞춘 수정 버전입니다. 워크플로 PNG의
편집 가능한 캐릭터 프롬프트 연결과 구형 코어 오류 안내를 포함합니다.
업데이트 후 ComfyUI를 재시작하고 브라우저를 새로고침하세요.

Verification

Local Python regression: 405 tests run, no failures, 7 conditional skips.
UI regression: 15 test files passed. The new release regression checks the
runtime version, manifest, all five README links and the release-note file.
This patch changes version metadata and distribution documentation; the existing
pose-image audits remain the generation-quality evidence.

1.0.9 Hangeul — Krea 2 / MOSS 3D Pose Support

1.0.9 Hangeul

2026-10-30. Experimental trained 3D pose guidance for Krea 2 Turbo and MOSS,
alongside the existing numbered character prompts and regions.

• RS 3D Pose Adapter (Experimental) receives MODEL, matching VAE and
  scene_json from RS 3D Pose Masks. Its MODEL feeds the standard RS sampler;
  the separate building pipe still supplies character ownership through Finalize.
• Krea 2 Turbo uses a trained OpenPose LoRA and cached pose-reference K/V.
  MOSS uses trained person-color/depth control and a replacement input embedder.
  Adapters download on first use, with progress and revision/checksum verification.
  Zero strength or an empty application interval downloads no weights.
• New sampler layout default auto chooses moss_native with active MOSS pose
  control, otherwise regional. Explicit saved choices remain unchanged.
  MOSS native converts masks to enclosing boxes while retaining learned pose
  maps; RS regional boundary controls do not apply in native mode.
• Existing Anima 2B LLLite setup and original 3D output slot indices are preserved.

Start here

한국어 가이드 ·
English setup ·
Genshin workflow PNGs

Restart ComfyUI and refresh its browser after updating. Select matching base
models/encoders/VAEs; install requirements-reference.txt with ComfyUI's Python
if optional dependencies are missing. First-use adapter downloads are about
229 MB (Krea) and 313 MB (MOSS); base models are not included.

Evidence and limitations

The 36-image audit uses Jean and Eula,
two models, six poses and two ON seeds, with one matched OFF/ON seed per scene.
No pose-action text or style/character LoRA is used; trained control LoRAs are
part of the adapters. Raised arms and sitting respond in both models. Exact
contact, orientation and joints remain unreliable; one MOSS side-view case swaps
character/pose roles. Prior extra-person failures with MOSS style LoRAs remain
unresolved. Execution success is not a pose-quality success rate.

Tested models: Krea 2 Turbo FP8 scaled and MOSS SFT BF16, current native core
5cc4b87. Other checkpoints, quantized pose patches and character-LoRA
combinations require independent tests. Image references, progressive sampling
and other attention replacements are restricted on this experimental path.
MOSS map start/end controls the conditioning interval; its control LoRA and
expanded input embedder remain installed throughout sampling.

Sources and redistribution terms are in third-party notices.
The full regression and installed-runtime release checks are recorded in the
release verification report.

한국어: Krea 2는 리저널＋전용 포즈 어댑터, MOSS는 네이티브 JSON＋색상·깊이
어댑터를 우선 사용합니다. 마스크만 전달하는 것과 학습된 포즈 유도는 다른 경로이며,
캐릭터 영역 파이프와 어댑터 MODEL을 모두 연결해야 합니다. 큰 자세를 유도하는
시험 기능으로 제공하며 정확한 손·접촉이나 캐릭터 배정을 보장하지 않습니다.

1.0.8 Hangeul — Faster Anima Regional References

Anima reference predictions now skip algebraically cancelled work while retaining independent character references, masks, strengths and configured duration. Existing workflows need no new setting.

In paired single-run RTX 5060 Ti measurements (1152×896, 30 steps, Euler ancestral / SGM uniform, CFG/reference guidance 4.5, global RSCV5 0.9, reference scale 0.8, references active throughout):

Condition	Before	1.0.8	Reduction
Two numbered regional v2 references	148.93 s	94.40 s	36.62%
One global v2 reference	71.95 s	45.24 s	37.13%

These are before/after comparisons within each scene, not a controlled comparison of global versus regional overhead. Model loading and PNG saving are excluded. Reduced-precision batching changes pixels; clothing and reference quality limitations remain. Unequal guidance/custom CFG extensions keep the previous calculation.

Validation: 387 Python tests passed, 7 conditional skips; all 15 UI files passed. A live installed-runtime smoke generated two independently referenced character regions with v2, recorded 12 skipped referenced-negative predictions across 6 steps, and released all per-sample reference cache bytes.

한국어: 영역별 참조를 그대로 유지하면서 중복 예측을 줄였습니다. 실측 조건에서 약 37% 빨라졌으며, 참조를 16스텝에서 끄는 방식이 아닙니다. 기존 워크플로와 어댑터 선택 방식은 유지됩니다. ComfyUI 재시작 후 브라우저를 새로고침해 주세요.

Full conditions and limitations: 1.0.8 notes.

1.0.7 Hangeul

1.0.7 Hangeul

2026-10-29. This release includes independently selectable experimental tools;
it does not enable all adapters together or replace existing sampling defaults.

3D scene and pose guidance

• Add RS 3D Pose Masks (Experimental): a local, offline editor with skinned
  human bodies, joint/hand/finger controls, person transforms and camera controls.
  Apply the scene to create numbered full/visible silhouettes in the existing
  building pipe. The browser and CPU renderer share the same scene/rig data.
• Output a separate shaded pose_image for a compatible learned pose adapter.
  Native Apply Anima LLLite patches can run alongside RS regional routing.
  Masks need no inference weights; learned pose guidance requires a separately
  installed matching adapter. Actual pose generation is validated on Anima 2B.
• Preserve legacy ellipsoid scenes, numbered assignments, output indices and
  existing building-pipe data. Clip masks at the canvas edge. Pose guidance is
  not exact joint/contact control; frontal multi-person seating remains weak.
• Include a real generated workflow PNG,
  setup instructions and feature-isolation results.

Reference alternatives

• Add explicit Auto: Anima v2 (2B) selection on the experimental MODEL
  reference adapter. It computes independent reference K/V banks once and reuses
  them during generation; numbered references retain their regional routing.
  Active use downloads pinned weights with hash verification and progress.
• Keep the pilot reference path as the automatic default. Full-duration v2 was
  about 29–34% faster in the measured two-character cases, with higher peak
  allocated memory and visible hair/clothing errors. These are small comparisons,
  not a universal benchmark or a recommendation to replace the pilot.
• Keep references active through the final step by default. Cutting at step 16
  remains a quality tradeoff, not a recommended optimization preset.
• Include the separately selectable experimental Role K/V backend and
  model-matched pilot adapters for Anima 2B / 2.9B. Each path has different
  restrictions: Role K/V and v2 target 2B; v2 rejects character LoRAs, whereas
  Role K/V requires its documented Normal-mode/balanced-routing setup.
• Pose patches and image-reference adapters cannot be combined in this release.
  No weights are bundled. See pilot setup,
  Role K/V, v2 results and
  2.9B validation.

Evidence and update

Legacy compatibility: actual CUDA attention tests passed with PyTorch 2.5.1,
xFormers 0.0.28.post3 and ComfyUI v0.3.76 (FP16/BF16; native and regional
paths). This is kernel/API verification, not a full image-quality benchmark.
Windows now rechecks reference-weight hashes on reuse to avoid timestamp-cache
collisions after a same-size file overwrite. Scope and reproduction.

The 3D editor integration has 12 paired generation outputs and one complete
UI-submitted workflow execution. A separate 24-image ablation uses three scenes,
two seeds and four conditions: baseline, regional only, pose only, both. Style
LoRAs, image references and action captions were absent in that ablation. Pose
guidance induced raised-arm and oblique seated poses; frontal two-person seating
failed even without RS regions. Exact poses and identity are not guaranteed.
The reference-v2 experiment contains 14 generated images; its speed/memory and
quality findings are documented separately from the pose results.

Restart ComfyUI and refresh the browser after updating. Use the existing sampler
and character workflow normally unless explicitly selecting these new options.
The 3D editor is in this local-model extension; no separate 3D editor is installed
by the NovelAI Request package. Upstream code, asset and model-license references
are retained in third-party notices.

한국어: 3D 인체 편집으로 번호별 마스크를 만들고, 별도 Anima 2B 포즈
어댑터로 자세를 유도합니다. 마스크만 사용하면 추가 가중치는 필요 없습니다.
두 기능은 독립적으로 쓸 수 있으며, 관절·손 접촉의 정확성은 보장하지 않습니다.
참조 v2는 별도 선택 기능으로 추가했고 기존 참조 기본값은 유지합니다.

Final validation

• Local Python: 391 tests ran, 384 passed and 7 optional tests skipped.
• JavaScript UI: all 15 test files passed.
• Isolated legacy CUDA: all 4 tests passed on PyTorch 2.5.1 / xFormers 0.0.28.post3 / ComfyUI v0.3.76, FP16 and BF16, RTX 4070 Ti SUPER. These test attention kernel/API compatibility, not full image quality.
• Component experiments, settings and limits are included in the optional validation ZIP. Extract and open the individual index.html galleries.
• ZIP SHA-256: b5a9d94f73cfa6f4d70b308ca9043409aa8137beb99b21a09701ddc833e2a274.

Downloads

The included workflow PNG can be dragged into ComfyUI. RS-MC-1.0.7-Validation.zip contains the optional comparison galleries, source images, measurements and CUDA logs; these are not downloaded by Git installation. Download model weights through the documented opt-in adapter setup.

1.0.6 Hangeul — Progressive Resolution Sampler

1.0.6 Hangeul

Released source update dated 2026-10-27.

• Add RS Progressive Resolution Sampler, a separate T2I node using the existing MODEL, CLIP, VAE and finalized character pipe. The existing RS MultiCharacter Sampler remains available.
• Choose final width/height, start scale 0.5–1.0 (default 0.67), sampler, scheduler, total steps, CFG and seed. Middle sizes and expansion steps are calculated automatically. No additional model or dependency is required.
• Expand the denoised latent with Lanczos and apply the model's native noise scaling. Recalculate regional masks at each size; below final resolution keep exclusive character-region separation at least 0.99. Preserve the existing overlap policy.
• Retain character LoRA Normal/Precision modes, masks, text regions, camera controls and outputs. Store the effective stage plan in PNG/WebP generation metadata and support restoring image settings.
• Add localized UI/help in English, Korean, Japanese and Simplified/Traditional Chinese, plus a workflow embedded in an actual generated example PNG.

Use and limits

See the progressive resolution guide and workflow PNG. For 1152×1728, 30 steps and scale 0.67, the usual plan is 768×1152 for 8 steps, 960×1440 for 8 steps, then 1152×1728 for 14 steps. Total steps are shared across the stages.

This trades some detail for shorter sampling at larger resolutions. Lower start scales can lose detail, alter composition or merge people; quality and speed gains are not guaranteed. Scale 1.0 uses ordinary fixed-resolution sampling. SD1.5 is supported for execution compatibility only, with no recommendation for high-resolution quality. Native MOSS keeps its own JSON layout behavior. Stateful solvers restart their history at expansion. Body-part guidance combined with text regions retains the existing unsupported-combination error.

Validation

The development candidate was verified in an isolated ComfyUI on RTX 4070 Ti SUPER: 34 matched quality images across Anima, SDXL, SD1.5, Krea 2, MOSS attention/native and FLUX.1/2/3. Additional short executions covered installed distilled/Dev/Schnell/SFT variants. These are limited comparisons, not every checkpoint or sampler/scheduler combination. A warmed scale-1.0 comparison was pixel-identical to the ordinary sampler in all eight branches. Actual generation also covered regional LoRAs, masks/text, body-part guidance without text, metadata restoration and WebP autosave.

The development regression run passed 275 Python tests, with 8 conditional skips, and all 15 JavaScript suites. Seven skips require the optional GGUF fixture; one applies only to older ComfyUI cores. The feature guide records model-specific quality limits, representative solver checks and references. The prior fixed-stage Lanczos/SPEED comparison is design evidence, not an independent speed benchmark of this final node.

References

SPEED ComfyUI implementation · research implementation · research description. RS compared frequency-preserving expansion, noise/time correction and expansion timing after its earlier Lanczos trials. The released node uses Lanczos; it does not contain SPEED's DCT expansion or frequency-derived automatic timing. See third-party notices.

한국어: 큰 이미지를 조금 더 빠르게 샘플링하는 대신 디테일을 일부 희생하는 별도 노드입니다. 시작 배율은 0.5~1.0이며 낮을수록 디테일 손실·구도 변화가 커질 수 있어 품질을 보장하지 못합니다. 기존 멀티캐릭터 샘플러와 캐릭터 파이프 기능은 그대로 사용할 수 있습니다. 업데이트 후 ComfyUI를 재시작하고 브라우저를 새로고침하세요.