Skip to content

Ecosystem digest 2026-07-18 #129

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
laboratoiresonore wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Ecosystem digest 2026-07-18 #129

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 68 additions & 0 deletions _dev_docs/ecosystem_digest_2026-07-18.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
# Ecosystem digest — 2026-07-18

_Generated by the every-48h cloud routine (first run). Runtime: ~10m. Backends reached: WebSearch (14 queries) + Hugging-Face MCP (4 queries). LAN-only backends (comfy_manager, local_llm) deferred to operator's local cron. `tools/upgrade_research.py` NOT present in current checkout — see Run metadata._

## Headline — top 5 upgrade candidates

| # | Spellcaster method/arch | Candidate | Source | Downloads/Rating | Why |
|---|---|---|---|---|---|
| 1 | flux checkpoint family | **FLUX.2 [klein] 4B** (`black-forest-labs/FLUX.2-klein-4B`) | HF | very active (multiple fp8/int8/GGUF forks in past 30d) | Apache-2.0, fits 13 GB VRAM (comfortable inside 16 GB ceiling), sub-1s inference, unifies text-to-image + multi-reference edit — direct replacement for older FLUX 1.x-dev pipelines |
| 2 | image edit / instruction edit | **Qwen-Image-2.0** (7B MMDiT, Alibaba, Feb 10 2026) | Web | #1 on AI Arena for both t2i and edit | Unified gen+edit in a single 7B model (down from 20B), native 2K, professional typography; strong candidate to displace separate Qwen-Image + Qwen-Image-Edit chains |
| 3 | ComfyUI runtime | **ComfyUI v0.28.0** (Jul 15 2026) | GitHub | latest of 8 minor releases ahead of bundled 0.20.3 | Int8/int4 Turing support, multi-threaded model loading, SeedVR2 upscaler, LoRA-apply memory improvements — baseline is 8 versions stale |
| 4 | face identity on FLUX | **PuLID-Flux2 v0.6.2** (`iFayens/ComfyUI-PuLID-Flux2`, Mar 21 2026) | GitHub | first working PuLID for FLUX.2 | Supports Klein 4B/9B + Dev; if you adopt candidate #1 you'll want this for face-lock replay |
| 5 | video (long-form) | **LTX-2.3** (Lightricks, Mar 26 2026) | Web | Apache-2.0, 22B | Native 4K @ 50 fps up to 20 s, sync audio in a single forward pass, native portrait mode — heavy (>13 GB, exclusive-load), but currently the only permissively-licensed high-fidelity local option in this class |

_Ranking notes: candidate #1 wins on VRAM fit + license + ecosystem momentum (dozens of quantizations shipped in the last 30 days). Candidate #5 is flagged **heavy** for the RTX 5060 Ti — treat as exclusive-load._

## Ecosystem releases since last digest

_First run; comparing against operator's memory baselines._

- **ComfyUI**: **v0.28.0** (Jul 15 2026) — baseline 0.20.3 is 8 minor versions behind. Intermediates worth noting: 0.21 (Gemma4 + auto-regressive video, IPEX dropped), 0.22 (MoGe, Stable Audio 3), 0.23 (multi-threaded model loading, MediaPipe/PixelDiT/Microsoft Lens support), 0.28 (int8/int4 Turing, SeedVR2 upscaler, Save3D nodes, ByteDance Seed Audio 1.0). https://github.com/Comfy-Org/ComfyUI/releases
- **krita-ai-diffusion**: **v1.52.1** (Jun 30 2026) — baseline v1.50 is 2 minor versions behind. v1.52.0 (Jun 28) adds Krea 2 support, extends regional prompting to Anima, fixes v1.51.x auto-updater bug. https://github.com/Acly/krita-ai-diffusion/releases
- **LM Studio**: **v0.4.18** (Jun 26 2026) — MTP speculative decoding shipped in 0.4.14, mlx-engine 1.8.1 with parallel predictions for Qwen 3.5/3.6 + Gemma 4 in 0.4.13, MCP OAuth in 0.4.10, Qwen 3.6 support in 0.4.12. https://lmstudio.ai/changelog
- **Models — substantively new since baseline:**
- **Flux 2 Klein** — released as `black-forest-labs/FLUX.2-klein-4B` + `FLUX.2-klein-9B` (Apache 2.0). Active downstream: `unsloth/FLUX.2-klein-9B-GGUF`, `6Morpheus6/FLUX.2-klein-4B-fp8-diffusers`, `SamuelTallet/FLUX.2-klein-4B-SDNQ-8bit-dynamic-hadamard256`, ConvRot int8 variants from `wraps/*`, all last-modified in the past week.
- **WAN**: 2.6 (Dec 2025, i2v with 1–9 refs, 720p/1080p, 3–15s clips) and **2.7 Image** (Apr 1 2026, Apache 2.0, adds first/last-frame control + 9-grid i2v + instruction edit + video recreate). Note: 2.7 weights primarily distributed via ModelScope; HF `Wan-AI` org still hosts 2.1/2.2 only as of this run.
- **LTX-Video**: **LTX-2.3** (Mar 26 2026), 22B, sync audio, 4K/50fps/20s, Apache 2.0. No 2.4/3.0 yet.
- **Z-Image**: **Z-Image base** released Jan 27 2026 by Alibaba Tongyi Lab (24 GB VRAM — heavy). Turbo remains 6B/fast tier. Z-Image-Edit / Z-Image-Omni-Base announced, not yet released.
- **Qwen-Image-Edit**: superseded by **Qwen-Image-2.0** (Feb 10 2026), 7B unified gen+edit, 2K native, currently #1 on AI Arena in both t2i and edit categories.

## HF trending shift

_Skipped — first run, no prior digest to diff against. Trending-score sort on `text-to-image` returned only legacy noise (2022 test repos) and was not useful this cycle; will re-attempt next run with a task-tag filter._

## Upgrade-cluster watchlist

| Cluster | Status this run |
|---|---|
| PuLID-Flux2 | **Active** — v0.6.2 (Mar 21 2026) supports Klein 4B/9B + Dev; InsightFace + EVA-CLIP stack. Training scripts temporarily removed pending stability fixes. |
| MagCache | **Steady** — official Wan 2.2 support now reports 1.5–2× speedup (down from 2–3× on Wan 2.1 / HunyuanVideo). Single-sample calibration still the differentiator. |
| HyperSwap | **Regression risk** — ReActor issue #226 (Apr 2026) reports "black-circle instead of face" on all `hyperswap_*_256.onnx` variants; inswapper/reswapper still work. Recommend pinning to inswapper128 until upstream fixes. |
| Differential Diffusion inpaint | **Quiet** — no material 2026 improvement over baseline soft-inpainting workflow. DiffSynth-Studio added audio-video inpainting on LTX-2.3, orthogonal to Spellcaster's image inpaint path. |

## Backend results

- **local_index**: SKIPPED (0 baseline candidates — `lora_calibrations_sfw.json` in current checkout is a skeleton with an empty `loras` dict; the populated calibration lives in the operator's local checkout only)
- **huggingface**: N/A — `tools/upgrade_research.py` not present. Substituted with 4 direct Hugging-Face MCP queries (see Section 4 in schedule)
- **civitai**: N/A — same reason. Substituted with WebSearch coverage of the watchlist clusters
- **comfy_manager**: SKIPPED (LAN-only)
- **local_llm**: SKIPPED (LAN-only)

## Action items for operator

1. **ComfyUI bundle bump** — 0.20.3 → 0.28.0 is 8 minor versions behind and covers real features (int8/int4 Turing, multi-thread model load, Save3D nodes). Priority: **high** — the drift will start blocking new model support.
2. **Wire FLUX.2 Klein 4B** — fits VRAM ceiling, Apache 2.0, downstream quantizations already stable. Pair with PuLID-Flux2 v0.6.2 for the face-lock replay path. Priority: **high** — the model of the moment for our size class.
3. **Evaluate Qwen-Image-2.0 as a Qwen-Image-Edit replacement** — 7B (vs prior 20B) with unified gen+edit could simplify the current dual-path chain. Priority: **medium** — worth a benchmark against existing edit workflows before committing.

## Run metadata

- **runtime**: ~10m (well inside 60m cap)
- **WebSearch call count**: 14
- **WebFetch call count**: 0 (search summaries were sufficient this cycle)
- **Hugging-Face MCP call count**: 4
- **Google-Drive MCP call count**: 1 (backup delivery)
- **Gmail MCP call count**: 0 — **server requires OAuth and this session is non-interactive**; operator can authorize in `claude mcp` / claude.ai connector settings, but for this run delivery falls back to git PR + Drive + PushNotification
- **spellcaster build_\* count**: 77 (baseline was 73 at 2026-05-13; +4 drift over ~2 months is normal — no anomaly)
- **upgrade_research.py**: **not present in checkout** — treated per section 10; synthesized digest from web + HF MCP data only. Recommend authoring/committing the script next cycle so real-index scoring returns.
- **next scheduled run**: 2026-07-20 ~16:00 UTC
Loading