civitai__civitai/docs/seo-ecosystem-review-checklist.md

SEO ecosystem pages — human review checklist

Dev-flagged claims on the ecosystem SEO pages that a human should verify against a source before we treat them as authoritative. Each links to the page (and the relevant section). Generated from the factCheck data in src/shared/constants/ecosystem-seo.constants.ts.

15 pages · 21 items. As of 2026-07-22 these were fact-checked against public sources (see per-item verdicts + source links). No claim was found to be outright wrong. The handful of open judgment calls were then resolved by dev decision on 2026-07-22:

• HappyHorse & Seedream arena rankings — keep both out of the evergreen pages (already excluded).
• Anima weighting — page now says weighting works (use higher values like :2), replacing the old "don't use weight syntax" tip.
• Specific video clip durations (5s/10s/8s, etc.) — removed from all video-page copy and example settings; resolution/fps retained.

Resolved items had their factCheck flag removed from ecosystem-seo.constants.ts, so a regenerated checklist would no longer list them.

Legend: ✅ verified accurate · ⚠️ verified, decision since resolved · ❌ would mean a correction is needed (none found).

---

Illustrious

Page: https://civitai.com/ecosystems/illustrious

• promptTips (open →) — booru-tag prompting guidance — Orchestrator prompt guide returned a generic fallback for this key — tips grounded in the model card + general SDXL/booru practice, not a model-specific guide.
  • ✅ Verified 2026-07-22: Correct. Illustrious is a Danbooru-trained SDXL anime fine-tune; comma-separated booru tags (important tags front-loaded), not natural-language sentences, is the established convention. Sources: Illustrious prompt guide, booru-style tagging guide.

NoobAI

Page: https://civitai.com/ecosystems/noobai

• promptTips (open →) — booru-tag prompting guidance — Orchestrator prompt guide returned a generic fallback for this key — tips grounded in the model card + general booru practice, not a model-specific guide.
  • ✅ Verified 2026-07-22: Correct. NoobAI is a Danbooru+e621-trained SDXL fine-tune; same booru-tag convention as Illustrious. Sources: Illustrious/NoobAI style explorer, booru-style tagging guide.

Pony

Page: https://civitai.com/ecosystems/pony

• promptTips (open →) — score_ tag prompt guidance — Orchestrator prompt guide returned a generic fallback for this key — tips grounded in the model card + general Pony/booru practice, not a model-specific guide.
  • ✅ Verified 2026-07-22: Correct. Pony Diffusion V6 XL's documented convention is leading the positive prompt with score tags (score_9, score_8_up, score_7_up, …). Sources: official Pony V6 XL model page, Pony prompt tags guide.

HiDream

Page: https://civitai.com/ecosystems/hidream

• overview (open →) — 17B sparse mixture-of-experts transformer — Verify the parameter count / MoE architecture against HiDream I1 release notes.
  • ✅ Verified 2026-07-22: Correct on both counts. The HiDream-I1 paper describes a 17B-parameter image foundation model built on a sparse Diffusion Transformer with a dynamic Mixture-of-Experts (token-routed FFN experts + a shared expert). Source: HiDream-I1 paper (arXiv 2505.22705).

Anima

Page: https://civitai.com/ecosystems/anima

• promptTips (open →) — weight-syntax guidance — The model card and the prompt guide CONFLICT on whether weight syntax works — the guide was followed. Confirm which is correct.
  • ✅ Resolved 2026-07-22: Not actually a conflict, it's an omission. The Hugging Face model card states prompt weighting works but needs higher-than-usual values (e.g. (chibi:2)); the orchestrator recipe just didn't mention it. The page previously told users not to use weight syntax — now corrected to "weighting works, use a higher weight like (word:2)." Flag removed from config. Sources: Anima HF card, Civitai orchestrator recipe.

Z-Image

Page: https://civitai.com/ecosystems/z-image

• overview (open →) — compact ~6B architecture — Verify the parameter count against the Z-Image release notes.
  • ✅ Verified 2026-07-22: Correct. Alibaba Tongyi Lab's Z-Image-Turbo (Nov 2025) is a 6B-parameter single-stream DiT (S3-DiT). Source: Tongyi-MAI/Z-Image-Turbo HF card.

Kling

Page: https://civitai.com/ecosystems/kling

• featuredExamples (open →) — 8-second example clip (imageId 133284218) — Duration is the real image meta value, but sits outside the guide's stated 5s/10s modes — verify it's correct.
  • ✅ Resolved 2026-07-22: The 8s was a real meta value and legitimate (Kling 3.0 supports flexible 3–15s durations), but per dev decision specific clip durations were removed from all video pages — settings now show model · resolution · fps only, and the "5s/10s modes" prose was made duration-agnostic. This removed the inconsistency entirely; flag removed from config. Sources: Kling length limits, Kling 3.0.
• comparison (open →) — peer facts (Seedance = ByteDance, Hailuo = MiniMax) + qualitative ratings — Peer positioning is editorial / general knowledge, not sourced metrics.
  • ✅ Verified 2026-07-22: Both attributions correct — Seedance = ByteDance, Hailuo = MiniMax. Qualitative ratings remain editorial by design. Sources: MiniMax/Hailuo, Hailuo.

Seedance

Page: https://civitai.com/ecosystems/seedance

• comparison (open →) — qualitative ratings (prompt adherence, native audio peers) — Editorial judgment, not sourced metrics — spot-check the peer cells.
  • ✅ Verified 2026-07-22: Factual cells hold — Seedance is ByteDance and generates native/synchronized audio (dialogue, ambient, foley in one pass). Prompt-adherence "edge" is relative (some comparisons rate Veo 3.1 higher), so keep it worded as a relative strength. Source: Seedance 2.0.

Grok Imagine

Page: https://civitai.com/ecosystems/grok

• promptTips (open →) — video motion/camera prompt tip — The real grok prompt guide is image-oriented (Aurora); there's no Grok-video guide, so the video-motion tip is general best practice, not sourced.
  • ⚠️ Verified 2026-07-22: Grok Imagine genuinely does image + video (image-to-video with motion prompts) and the image engine is xAI's Aurora — both confirmed. xAI describes prompting "the motion" incl. camera moves, but publishes no detailed prompt spec, so the specific camera-motion tips are community best practice, not officially documented. Fine to keep; just know it's not officially sourced. Sources: xAI Grok Imagine 1.5, Aurora launch.
• comparison (open →) — peer positioning / ratings — Editorial, not sourced metrics.
  • ✅ Verified 2026-07-22: Editorial by design; no factual error found in the peer positioning. (Aurora / dual-modality basis confirmed above.)

HappyHorse

Page: https://civitai.com/ecosystems/happyhorse

• attribution (open →) — "attributed to Alibaba" — Corporate parent unconfirmed — the guide said "Alibaba, via fal.ai"; Civitai groups it under an Alibaba–Taotian family. Confirm the real owner.
  • ✅ Verified 2026-07-22: Correct. HappyHorse-1.0 is built by Alibaba (Taotian / "ATH" business group) with fal.ai as the official API partner (not the developer). Artificial Analysis lists the maker as "Alibaba-ATH" and notes the specific lab attribution is not independently verified, but Alibaba as the owning company is consistent. Sources: fal HappyHorse, fal press release, Artificial Analysis.
• overview (open →) — native synchronized audio + physics-aware motion — The model card leads with an unverified "#1 on the Artificial Analysis Video Arena / Elo 1416" claim, deliberately excluded here — decide whether to include.
  • ✅ Resolved 2026-07-22: Audio + physics claims confirmed (fal documents single-pass joint audio-video with native lip-sync + physics realism). Dev decision: keep the ranking out (it stays excluded). For the record, the #1 ranking is real but the "Elo 1416" figure was wrong anyway (fal lists 1333 T2V / 1392 I2V). Flag removed from config. Source: fal HappyHorse.

Nano Banana

Page: https://civitai.com/ecosystems/nano-banana

• promptTips (open →) — editing / prompting tips — No model-specific prompt guide exists (generic fallback) — tips grounded in the model card's documented editing capabilities, not a guide.
  • ✅ Verified 2026-07-22: Correct. Nano Banana = Google's Gemini 2.5 Flash Image, marketed around natural-language targeted editing (backgrounds/objects/colors/textures) and strong character consistency — exactly what the tips describe. Sources: Google Developers blog, Vertex AI.

Seedream

Page: https://civitai.com/ecosystems/seedream

• promptTips (open →) — prompting tips — Prompt guide was a generic fallback — tips grounded in the model card. ByteDance publishes an official Seedream guide worth mirroring.
  • ✅ Verified 2026-07-22: Grounding is sound; ByteDance's Seed team does publish official Seedream material worth mirroring. Source: ByteDance Seed — Seedream 4.0.
• overview (open →) — native 2K / 4K resolution — Sourced from the model card. The card's original "#1 on the Image Arena" claim was removed as unverified — decide whether to include.
  • ✅ Resolved 2026-07-22: 2K/4K is native (generated, not upscaled — Seedream 4.0; 3.0 capped at 2K). On the arena claim: it was genuinely #1 on both Artificial Analysis text-to-image and editing arenas at the Sept 2025 launch, but has since been overtaken (GPT Image 2 debuted #1; Seedream slid to ~#7). Dev decision: keep the ranking out (already excluded). Flag removed from config. Sources: ByteDance Seed, Artificial Analysis model page.

Veo 3

Page: https://civitai.com/ecosystems/veo-3

• comparison (open →) — peer native-audio cells (Sora 2 = Yes, Kling = No, …) — Cross-checked against the Sora 2 config (consistent), but peer cells are general knowledge, not re-verified against each provider.
  • ✅ Verified 2026-07-22: All cells accurate. Veo 3 generates native audio (dialogue/SFX/ambient); Sora 2 = Yes (native audio since its Sept 2025 launch); Kling = No (no native synchronized audio). Sources: Google DeepMind Veo, OpenAI Sora 2.

Sora 2

Page: https://civitai.com/ecosystems/sora-2

• overview (open →) — synchronized audio (speech, SFX, soundscapes) — Claimed from the model card only; the sora2 prompt guide does not mention audio. Verify the capability.
  • ✅ Verified 2026-07-22: Confirmed. OpenAI states Sora 2 generates video with natively synchronized audio incl. dialogue/speech, sound effects, and ambient soundscapes. Source: OpenAI Sora 2.
• comparison (open →) — peer native-audio / provider cells — From general knowledge / sibling configs, not independently re-verified.
  • ✅ Verified 2026-07-22: Consistent with the Veo 3 check above — peer native-audio/provider cells hold. Sources: OpenAI Sora 2, Google DeepMind Veo.

Chroma

Page: https://civitai.com/ecosystems/chroma

• localRun (open →) — ~12GB+ VRAM / ~9–18GB weights — Estimates, not stated in the model card (card only confirms FP8/GGUF variants exist). Verify against Chroma release notes.
  • ✅ Verified 2026-07-22: Accurate. Chroma (8.9B params, based on FLUX.1-schnell, Apache 2.0) runs on ~12GB+ VRAM via FP8/Q8 quants; weights span ~9GB (FP8/pruned) to ~17–18GB (FP16) — the "~9–18GB" range is right. Sources: Chroma HF README, willitrunai Chroma.
• comparison (open →) — peer parameter counts (SDXL ~3.5B, etc.) — Approximate — adjust if you want exact figures.
  • ✅ Verified 2026-07-22: SDXL ~3.5B is correct (2.6B UNet + ~817M dual text encoders ≈ 3.4–3.5B total, per the SDXL paper). Source: SDXL paper (arXiv 2307.01952).

---

Second pass — unflagged prose (2026-07-22)

Beyond the dev-flagged claims above, the descriptive/architecture/version prose on every page was fact-checked against primary sources (papers, HF cards, official release notes). ~85 specific claims across all 24 pages. Most held up; the corrections below were applied to ecosystem-seo.constants.ts.

Corrected (were inaccurate)

• Imagen 4 — negative-prompt field. Page told users to use "the separate negative-prompt field" (3 places). ❌ negativePrompt was removed as of Imagen 3; Imagen 4 has none. Fixed: rewrote all three to "phrase everything positively — no negative-prompt field." Source: google-genai #339.
• Seedream — 4K debut version. Page said "Seedream 4.5 adds 4K" (4 places). ❌ 4K arrived in Seedream 4.0; 4.5's upgrade was text rendering. Fixed: 4.5 → 4.0 everywhere. Source: ByteDance Seed — Seedream 4.0.
• Flux.2 — "Max" tier. Page listed API tiers "Pro, Max, and Flex." ❌ Max is not a documented Flux.2 tier (only Pro/Flex are API-only; Dev/Klein open). Fixed: dropped Max (now "Pro and Flex"). Source: BFL Flux.2 blog.
• Flux.2 — hex color codes. Claimed it "accepts hex color codes tied to specific objects" (overview + a promptTip). ⚠️ No primary source (not in BFL blog or GitHub README). Fixed: replaced with "name exact shades and materials."
• Chroma — CFG default. PromptTip said "default guidance scale is around 5.0." ⚠️ The official HF example uses ~3.0; no source for a 5.0 default. Fixed: now "official example uses ~3.0." Source: Chroma HF.
• Krea 2 — "Large ≈ 2× Medium parameters." ⚠️ Krea publishes no per-tier parameter counts. Fixed: dropped the 2× claim (kept the softer-post-training distinction). Source: Krea 2 docs.
• Veo 3 — "returns a generic PG clip (no refund)." ⚠️ Real behavior is safety-filtering, but the specific "generic PG clip / no refund" framing isn't documented by Google (3 places). Fixed: softened to "explicit prompts are filtered" while keeping the SFW steer.
• Seedance — resolution. Page said "480p or 720p native." ❌ Understated — Seedance 2.0 supports 480p / 720p / 1080p / 4K (Fast capped at 720p). Fixed: "from 480p up to 4K native" (2 places). Source: Scenario — Seedance.

Verified accurate (spot-check highlights, no change)

• Flux.1 12B rectified-flow transformer, Dev 512 / Schnell 256 tokens, Dev/Schnell/Krea/Kontext variants — all ✅. (Minor: encoder is CLIP+T5, page simplifies to "T5"; left as a fair simplification.)
• Flux.2 Mistral Small 3.2 encoder, Klein 9B/4B — ✅.
• Chroma de-distilled Flux.1-schnell ~8.9B, T5, Apache-2.0, Base/HD/Flash/Radiance variants — ✅.
• SDXL 1024px, base+refiner, dual CLIP (77-tok), weighting math — ✅. Illustrious=OnomaAI, NoobAI=Laxhar Lab (+e621), Pony=AstraliteHeart (clip skip 2, Euler a/25 steps) — ✅. Qwen-Image 20B, Qwen-Image-Edit — ✅ (its 16GB/20GB local figure is the FP8 path; full precision is ~40GB — acceptable as the practical number).
• Anima 2B, CircleStone×Comfy Org, Qwen 3 0.6B encoder + Qwen-Image VAE, Sept-2025 cutoff, Base/Aesthetic/Turbo — ✅.
• Z-Image ~6B S³-DiT, Turbo 8 steps, fits 16GB / sub-second on H800 — ✅.
• Wan Wan-VAE, 1.3B–14B, 2.1 first open CN+EN in-clip text, 2.2 MoE + 5B TI2V, 2.7 is current — ✅. LTXV DiT/real-time/T5, LTX-2 (19B) joint A/V, LTX-2.3, Sulphur 2 — ✅.
• Kling Kuaishou, 3.0 current (Feb 2026) — ✅. Grok Aurora autoregressive image model + dual image/video — ✅. HappyHorse unified-Transformer native A/V, EN/CN/JP, v1.1 — ✅ (developer is an ex-Alibaba/ex-Kuaishou spin-off, which supports keeping the hedged "attributed to Alibaba" wording + its existing flag).
• Nano Banana = Gemini 2.5 Flash Image; original/Pro/2 versions real — ✅. Sora 2 OpenAI, successor to Sora (Feb 2024), synchronized audio — ✅.

Freshness disclosure

A standing note — "AI models move fast — new versions ship often, and a model's capabilities or Buzz cost can change. For the latest, check the model's own page before you generate." — now renders on every video and API-only page (gated on isVideo || apiOnly), so version-drift is disclosed rather than asserted.