This isn't an article's worth of material—it's about twenty links: a Quark netdisk share, six GitHub repositories, ten trainer images, two registration pages, one redemption code. The person who posted them is T8mars, a creator in the ComfyUI scene known for the "Zhenzhen" integration packs and an API proxy site.
Cross-checking against GitHub and several sites, this list is more structured than it looks. It is essentially the entire supply chain of AI image and video creation, compressed by one person into a netdisk link and a set of registration pages.
Nine Days: From Model Release to Full Toolchain
Start with the timeline—the sharpest part of this list.
On September 20, Alibaba's Qwen team open-sourced Qwen-Image-2.1: a 7B-parameter MMDiT architecture that puts text-to-image and image editing into a single model, with native support for transparency via Alpha channels and 2K output, under the Apache 2.0 license. ComfyUI shipped native support the same day.
Then came the community's speed. On September 21, T8mars's prompt-rewriting node hit GitHub—one day after the model's release. On September 23, the BlockCache acceleration node followed. On September 22, his API site added qwen-image-global-2.1 calls. On September 29, a multi-angle node was created. Within seven days, four tools—acceleration, prompt rewriting, multi-angle generation, and API access—grew around the new open-source model, plus companion workflow JSONs and integration packs.
I pulled data from all six repositories today: the veteran Comfyui-zhenzhen node has 751 stars; the open-source canvas T8-penguin-canvas has 606 stars (MIT license, 125 forks); three Qwen nodes created in September sit between 35 and 53 stars. Every repo was last pushed between October 2 and 6—still updating this morning. This is not a repo graveyard; it's a live production line.
A Third-Layer Cache That Saves Compute
The BlockCache repo deserves its own section. What it does can be said in one sentence: when a diffusion model generates an image, the computation passes through dozens of transformer blocks; if a layer's output is nearly identical to the previous step, that layer can be skipped and the previous result reused. The node checks "first-block change" and, if stable, skips the residual computation of subsequent blocks.
Anyone familiar with local deployment topics will recognize an old friend: last month's writeup on Strata covered three-tier placement of MoE expert weights—hot experts in VRAM, cold experts in RAM and SSD—with the core logic that "each token only needs a small subset of experts; idle weights don't need to be fast, just stored somewhere." BlockCache is the same law projected onto diffusion models—don't recompute what's been computed—only the caching target changed from weights to block residuals. One law, two battlefields.
What's more notable is how its README is written. First, it explicitly warns "don't stack gains from multiple backends." Second, it puts the cost of acceleration in plain sight: "the higher the threshold, the easier it is to skip computation, and the more likely you'll alter characters, text, and details." Third—and I think the rarest—it admits its own experimental mode lost: T8 Sage's hybrid routing "tested locally, was not faster than using Kitchen alone." A feature they built themselves, tested slower, written as slower—and they advise users to "enable one node alone first and compare against unaccelerated output."
Printing negative results in your own documentation isn't mainstream in open-source tooling. It's the same species as that recent piece about "contracts don't count, refusals count": speed claims aren't vouched for by the author; they're settled by users' local benchmarks.
A Two-Track Model Economy
The models in the list split into two kinds, each with its own playbook.
Qwen Image 2.1 has open weights, runs locally, and the ecosystem earns its keep in the tool layer—acceleration nodes save your GPU time, prompt rewriting saves your tuning time.
Seedance 2.0/2.5 and Seedream 5.0 are ByteDance's closed models, API-only. Without weights, the community's move is distribution: low-price API proxy sites, ComfyUI nodes calling them. The README lays this business out candidly: "non-profit site, keeping only a 5–10% website maintenance fee," "gross margin under 10%," "no room for negotiation, and no support for commercial secondary development—fan service only," "invoice quota is limited each month, and all taxes are borne by yourself."
That's the concrete form of "usage must follow the rules." Not pleasantries—the survival boundary of grassroots API distribution: stick to fan service, compress margins below ten percent, explicitly ban commercial secondary development—commercial use requires the model owner's license; that's a red line. The limited-invoice detail is especially real: the site's volume has grown large enough to worry about tax paperwork.
One ecosystem, two model economies coexisting: open models keep tool developers busy, closed APIs give distributors a spread. Neither has swallowed the other.
The Full Picture of the List
Sort the twenty-odd links by function and the structure emerges.
Compute layer: Chenyu Zhiyun (morning-feather cloud), exclusive GPU rental billed by the minute, with ten pre-installed trainer images—Qwen Image 2.1, Wan2.2, Flux.2 Klein, LTX2.3, MiniMax H3—spanning image, video, and audio-visual sync. Users who want to train LoRAs don't need to set up environments: rent a card, open an image, start training. Above that sits Alibaba Cloud Wuyi's cloud image, with a redemption code worth 200 core-hours.
Distribution layer: integration packs on Quark netdisk—the T8 CF pack is at V21, infinite canvas at v3.1.8, the Zhenzhen pack at V4.9C. Version numbers alone show the iteration speed.
Knowledge layer: OpenZhenZhen, an independent prompt gallery launched last week, freely sharing reusable generation recipes—e.g., "real-estate photos to 3D walkthrough video." Prompts here are published as recipes, not private assets.
Monetization layer: the API proxy's registration link carries a referral code; the cloud platform has redemption codes, Aix has invite codes, RunningHub has discount slots—the ecosystem's self-sustaining loop runs on these distribution codes.
Position on the Main Thread, and One Prediction
I've covered deployment collapse several times: on-device engines, unified-memory mini PCs, on-demand inference, libraries collapsing into byproducts of agent sessions. This list adds the consumer-creation side's full-stack sample—when models, tools, compute, knowledge, and distribution all collapse into the form of "netdisk link + registration page," the only remaining threshold is one thing: someone running errands for someone else. Integration packs install environments for you, images configure dependencies for you, proxy sites queue for you, galleries step on the pitfalls for you. Every step is a person saving another person's time.
A prediction that can be reconciled later: the middleman layer's margins will keep thinning. Its profit comes from exchange-rate spreads, queueing premiums, and compliance-risk subsidies; once official APIs cut prices directly or open more payment channels, that ten percent margin can't hold. If ByteDance's official prices drop another tier within six months, watch how these "under 10% margin" small sites transform—either toward services (integration packs, tutorials, custom workflows) or shut down.
End on a number. From model open-source to community full-toolchain coverage, Qwen Image 2.1 took nine days. In that open-source week, the official repo's star count and the community nodes' creation timestamps recorded this ecosystem's reaction speed more honestly than any press release.
---
*Verification notes: six GitHub repositories pulled via API (stars/forks/created/pushed/license, as of 10-06); Qwen-Image-2.1 release time and specs confirmed across multiple sources (2026-09-20 open-source, 7B MMDiT, unified text-to-image + editing, native RGBA); original READMEs of BlockCache-T8 and Comfyui-zhenzhen (skip-layer caching mechanism, dual-segment thresholds, hybrid routing admitting it lost, API site margin and rules self-descriptions, 10-02 changelog including the dola-seedream-5.0-flash series); Chenyu Zhiyun described by its official site as a per-minute GPU rental platform; OpenZhenZhen's independent-site form (crawlable gallery docs). Netdisk integration packs and trainer image contents were not tested (no account); related descriptions rely on source material and platform self-descriptions. Affiliate links and redemption codes are part of the creator's distribution structure, recorded as such and not a recommendation.*