Ecosystem digest 2026-07-18

[laboratoiresonore]

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 Install error #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
• 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

---

Generated by Claude Code