RC4_EXTERNAL_LANDSCAPE_2026-08-08.md

pipelines/RC4_EXTERNAL_LANDSCAPE_2026-08-08.md

RC4 — THE EXTERNAL LANDSCAPE ON CHARACTER LIKENESS (Aug 2026)

Lane: RC4 of the WHY-SO-FAR-OFF post-mortem. READ-ONLY; nothing applied.

Research date 2026-08-08. Every claim below is tagged VERIFIED (fetched live, URL inline),

REPORTED (secondary source, named), or UNVERIFIED (could not confirm at source — say so).

---

DERIVATION

DERIVED FROM:
  - docs/HANDOFF_2026-08-05_ACCOUNT_SWITCH.md §8.14 (L2811-2842) — the charter, Josh's
    verbatim grade, and RC4's exact wording: "EVALUATE NVIDIA-CLASS 3D (Edify 3D via
    partner access; Cosmos is scene/video not character) + any Aug-2026 leader on
    CHARACTER LIKENESS specifically - the bake-off scored symmetry/topology, never
    likeness-to-exemplar." (L2833-2835)
  - docs/pipeline_review/tech_research/PIPE_CHARACTER_MODELS_2026-08-06.md — read IN FULL.
    This lane's job is to extend/contest it, not repeat it. Load-bearing lines used:
      §3.1 "No open-weight general 3D generator's own gallery shows a photoreal human
        face... That silence is the finding, not an absence of evidence."
      §4 "Edify 3D - quad meshes with PBR, architecturally right, weights never released
        and the hosted NIM preview retired."
      §9 the Hunyuan territorial disqualification + the recorded lesson that the FIRST
        licence read was incomplete.
      §11 "Nothing in §2's bench is enrolled in a blind join" / Pixal3D+TRELLIS.2 never run.
  - docs/HANDOFF §8.8 (L431-433, L456) — the char-survey lane landing: "TripoSG holds,
    Hunyuan licence-killed".

CONTRACT ITEM 2 NOTE (required by the standing subagent contract):
  harness/route.py carries NO work-kind for an external-model post-mortem — this lane
  produces no canon and realizes no canon node. I therefore ran no route.py read-set.
  What I read instead is named above. No canon node was substituted by a proposal.

NOT DERIVED (authored judgment, and why it had no canon home):
  - The entire ranking in §7. Canon specifies the character, never the generator vendor.
    This is an engineering read of the public record as of 2026-08-08.
  - The claim in §1 that PIPE_CHARACTER_MODELS §3.1 searched the wrong family. That is my
    finding against a proposal-tier doc, which the contract permits (proposals are
    subordinate); it is grounded in named papers, not asserted.

---

0. THE HEADLINE — four findings

1. §3.1's class finding was derived from the wrong search. The prior survey looked at

*general* image-to-3D galleries and concluded the field does not do characters. A

character-specific sub-field exists, with its own benchmark protocol that scores

facial likeness to the reference directly. We benched the wrong family. (§1)

2. Seed3D 1.0 (ByteDance) is a total omission from the prior survey and is the single

highest-expected-value open candidate — 1.5B params beating Hunyuan3D-2.1's 3B on

geometry, full PBR, and the vendor's own character/face example. Its licence is

UNVERIFIED and that is a blocking read, not a footnote. (§2)

3. **NVIDIA is confirmed closed for this use, and Josh's "open world generation one" is

the wrong tool class** — Cosmos 3 generates video of scenes, not character meshes;

Edify's NIM preview was retired 2025-06-06 and the only live path is partner-mediated

through Shutterstock. The prior survey's §4 holds and hardens. (§3)

4. **The one independent likeness-specific head-to-head in the public record ranks closed

SaaS above everything we run** — and flags TRELLIS 2 for *broken hair strands*, which

independently corroborates our own hair finding. (§4)

The honest counterweight, stated up front: not one paper or gallery in this entire

sweep shows a photoreal human face. DreamCharacter-1's gallery is stylized. So the

prior survey's *deeper* point survives even though its stated reason is wrong: a photoreal

child face is unevidenced across the whole field. Nothing below promises to close that.

---

1. THE CHARACTER-SPECIFIC SUB-FIELD THE PRIOR SURVEY MISSED

§3.1 concluded from silence across TRELLIS.2 / Pixal3D / Step1X-3D / LATO.2 / AniGen /

Hunyuan3D-2.1 — all general object generators. Searching on *character likeness*

instead of *image-to-3D* returns a different literature that evaluates exactly our axis.

DreamCharacter-1 Table 1 (geometry, ULIP / Uni3D) — VERIFIED from arxiv.org/html/2607.07817:

Texture (Table 2): DreamCharacter-1 leads SSIM 0.9349 / LPIPS 0.0686 / FID 0.0319 /

CLIP-Sim 0.9576. No commercial baselines (Tripo/Rodin/Meshy) were compared.

Two things this table buys us even though the model is unavailable:

• **TRELLIS-2.0 is the strongest *open* model on the character axis** — above Pixal3D and

above Hunyuan3D-2.1. The prior survey listed TRELLIS.2 as an "untested upgrade lead"

(§11); this is published evidence that the lead is real *on characters specifically*.

It also directly supports RC3's charted first test.

• Pixal3D's Toys4K win over TripoSG does not transfer cleanly to characters — on this

character table Pixal3D sits *below* TRELLIS-2.0, inverting the §3.2 ranking.

The benchmark protocol is the reusable asset (RC5's answer). A character benchmark in

this line evaluates 100 character reference images on dimensions general benchmarks do not

score — **facial likeness, anatomical hand correctness, and full-body fidelity to the

reference — using VLM- and CLIP-based protocols, and for the face it selects a

VLM-verified camera distance that keeps the full face visible, then evaluates frontal and

profile views** (REPORTED, via search synthesis; the exact source paper was not pinned —

see §8 OWED). That camera rule is precisely the thing our turntable lacked. **RC5 should

copy this protocol rather than invent a likeness lens from scratch.**

---

2. SEED3D — THE OMISSION THAT MATTERS MOST

Absent from PIPE_CHARACTER_MODELS entirely. It is the backbone of the current SOTA

character paper.

Seed3D 1.0 (ByteDance Seed) — VERIFIED from seed.bytedance.com blog:

• Outputs **explicit watertight mesh + UV-mapped textures + PBR materials: "albedo,

metalness, and roughness"**.

• 1.5B parameters beats Hunyuan3D-2.1's 3B on geometry with "superior results" in

detail preservation. Texture/material claimed SOTA. Human eval, 14 evaluators,

"consistently higher ratings across all dimensions."

• It has a character example, and the claim is a likeness claim: the blog shows a

character generation example (facial features and fabric textures) where Seed3D 1.0

"accurately reconstructs facial features" while baselines "tend to lose reference

consistency." That is the one vendor face-likeness claim I found in the open-weight tier.

• Repo exists at github.com/Seed3D/Seed3D. arXiv 2510.19944.
• LICENCE: UNVERIFIED. Aggregator sources state "fully open source with Apache-2.0";

ByteDance-Seed's house licence is Apache-2.0 across Seed-OSS, SeedVR2 and Seed2.0. But

I could not confirm at file level — raw.githubusercontent.com/Seed3D/Seed3D/main/LICENSE

404s (branch name likely differs), the HF model card 401s, and the GitHub HTML render

came back incomplete. This is exactly the §9 Hunyuan trap — that lane read §6(d),

concluded clean, queued a 14 GB download, and found §5(c) afterwards. Do not spend a GPU

hour on Seed3D before the LICENSE file is in hand and pinned to docs/licence_records/.

Seed3D 2.0 — released 2026-04-23, API-only on Volcano Engine. Coarse-to-fine

geometry, unified PBR material model, MoE texture detail, VLM material priors,

part-level generation, articulated assets, scene composition. Blind human eval texture

preference >69%. Part-level + articulated is architecturally the answer to §3.1's

"one fused implicit surface" complaint — but there are no weights, so it is an API lane

only, and no character/face evidence was found for 2.0 specifically.

---

3. NVIDIA — THE CHARTER QUESTION, ANSWERED

Cosmos is confirmed scene/video, not character. Cosmos 3 launched 2026-05-31 (GTC

Taipei keynote at Computex): an open physical-AI omnimodel that generates text, images,

video and ambient sound plus robot action signals; 16B Nano and 65B Super. It is trained on

physical scenes — factories, warehouses, roads, manipulation workspaces — to produce

physically plausible video. It emits no character mesh. For character *motion* NVIDIA

points at ARDY (SIGGRAPH 2026, text→3D human/humanoid motion, real time), which the

prior survey already has under Apache code + Open Model License weights.

→ **Josh's "nvidias open world generation one applied to this" names a real and impressive

model that is the wrong tool class for a character asset.** That should be said plainly and

without hedging in the post-mortem; it is not a dodge, it is the answer.

Edify 3D — the partner access path, precisely. VERIFIED: NVIDIA's own editor's note,

dated 2025-06-06, reads that Edify is **no longer available as an NVIDIA NIM

microservice preview**, directing users to build.nvidia.com for other visual AI models.

No weights were ever released. The surviving path is partner-mediated, not NVIDIA-direct:

• Shutterstock's Generative 3D API, built on the Edify architecture, trained

exclusively on Shutterstock content including >500K ethically-sourced 3D models —

the "commercially safe / ethical" positioning is the product.

• Getty Images offers Edify-based custom fine-tuning on its commercially-safe service.

Read for us: the prior survey's §4 verdict stands unchanged — *NVIDIA has no local,

commercially-licensed, weights-downloadable character mesh generator.* The Edify path is a

third-party stock-media API with no local run, no character-likeness evidence, and a

training corpus of stock 3D objects. It is not a likeness route. RC4 closes NVIDIA.

---

4. THE ONE INDEPENDENT LIKENESS TEST IN THE PUBLIC RECORD

Vitalify Asia, July 2026 — seven image-to-3D tools, the same portrait image,

default settings, no post-processing. This is the only public head-to-head I found that

judges *facial similarity to a reference portrait* rather than object IoU.

Overall: Prism 3.1 for realism; Tripo P1 best speed/quality balance, called

"especially suitable for game development workflows."

How much weight this carries: n=1 per tool, sighted, single portrait, published by a

dev-services company — by our own §2.1 standard this is a LOOK, not a scored arm, and

it is enough to justify a probe, not to ratify a swap. Two things still make it the most

useful external datum in this report:

1. It is the only source that scores the axis Josh graded us on.

2. Its TRELLIS 2 hair finding — *broken strands* — was reached independently of us and

agrees with our own measured hair result (§2 of the prior survey found Hunyuan

resolves hair as flat sheets while TripoSG resolves strands). Two independent observers

landing on hair as the failure axis is worth more than either alone.

Note on identity: "Prism 3.1" is a Tripo model surfaced under 3D AI Studio's naming

(35 credits/generation there). Treat Prism 3.1 and Tripo P1 as the same vendor family.

---

5. THE CLOSED-SaaS TIER, ON LIKENESS-RELEVANT SPECIFICS

Standing Josh ruling keeps these reference-tier. Recorded because RC4 was asked to rank

them and because one of them is architecturally the closest fit to our actual problem.

Rodin Gen-2.5 (Hyper3D) — the best *architectural* match in the entire sweep:

• Accepts up to 5 multi-view reference shots. We have multi-view V3 plates. Every open

model in the prior survey is single-image conditioned. This alone is a structural

advantage on likeness.

• Quad topology selectable 4K–50K, or triangle up to 500K. Our canon LOD0 budget is

24k [CHAR_0001_CANON_SHEET.md L192-198]. A quad output inside that budget removes

the entire retopology stage that §3.1 named as the class blocker — the "~2 M-triangle

marching-cubes shell with no edge loops" problem simply does not occur.

• 10M+ polygon ceiling, 3D-native textures, texture delighting (our de-light stage),

symmetry forcing, micro detail control, HD texture enhancement. Exports GLB/FBX/OBJ/USDZ.

• Vendor claims ultra-fine surface detail "like fabrics, skin, and engravings."
• Commercial licence: NOT STATED on the site. ToS read owed before any output ships.
• Caveat from §4: the independent test flagged Rodin 2.5 for clothing-information loss.

Tripo P1 / Prism 3.1 — native 3D diffusion, engine-ready, clean low-poly topology,

full PBR set (albedo, normal, metallic, roughness). Licence: **Free plan grants NO

commercial use**; paid plans (Pro $19.90/mo, 3000 credits) grant broad rights to use,

distribute and derive revenue from outputs. Explicit, readable, and adequate for us.

CSM (Common Sense Machines) — strong on organic shapes and characters; all paid plans

include commercial rights; API access. Acquired by Google, January 2026 (REPORTED) —

worth noting beside the prior survey's other Google datum, GNM Head published 2026-07-28.

Sparc3D → Hitem3D — commercialized by Math Magic; Sparcubes/Sparconv-VAE at 1024³,

Hitem3D "1536p Pro". Possible correction to the prior survey, which recorded that

Sparc3D's weights were never released and the demo was pulled: a source states the GitHub

repo (lizhihao6/Sparc3D) ships pretrained weights, and a HF Space exists. **Low confidence,

UNVERIFIED** — flagged for a five-minute check, not asserted.

Hunyuan — corroborated, no change. 2.5 / 3.0 / 3.1 are hosted-only and were never

open-sourced; 2.1 (June 2025) remains the last open release. The prior survey's §9

territorial disqualification of 2.1 is therefore still the whole story, and there is no

newer open Hunyuan to reconsider.

---

6. LOCAL-RUN FEASIBILITY ON THE RTX 5090 32 GB

Standing hazard that applies to every open candidate here: sm_120 is untested for all

of them, and the prior survey §5 recorded gsplat PR #1034 (5090 support) still open with an

sm_120 illegal-memory-access issue filed 2026-08-04. Budget for a build fight.

---

7. THE RANKED SHORTLIST — for matching exemplary 2D character plates in 3D

1. Seed3D 1.0 — highest expected value, GATED ON A LICENCE READ. Only open-weight

candidate with a vendor face-likeness claim; full PBR; smallest model; and it is the

backbone the SOTA character paper chose to build on. If the licence is Apache-2.0 as

the secondary sources say, this is the model to bench against TripoSG immediately.

2. Rodin Gen-2.5 — best architectural fit, closed. Multi-view input matches our plate

stack; quad topology at 4K–50K lands inside the 24k LOD0 budget and deletes the

retopology stage that §3.1 identified as the class blocker; texture delighting is a

stage we already run. Buy one generation and look.

3. **Tripo P1 / Prism 3.1 — the measured likeness leaders, closed, licence-clean on a paid

tier.** Cheapest possible empirical answer to "is our chain the problem, or is 3D just

not there yet."

4. **TRELLIS.2 full PBR — already charted as RC3's first test, and now externally

corroborated** as the strongest *open* model on the character axis (0.8329/0.8261,

above Pixal3D and Hunyuan3D-2.1). Run it. Expect hair to be its weak axis.

5. IDOL (MIT) — reference tier for the photoreal-human/body half; animatable output.

6. DreamCharacter-1 / StdGEN / Seed3D 2.0 — architecture proof-points only. They show

the field's answer is post-training + semantic decomposition + part-level generation,

none of which we can download today.

DISQUALIFIED, with the clause:

• StdGEN checkpoints — "released checkpoints are for research purposes only."
• Hunyuan 2.1 / 2.5 / 3.0 / 3.1 — prior survey §9 territory clause; newer versions

never opened.

• Edify 3D — no weights; NIM preview retired 2025-06-06.
• Cosmos 3 — not a disqualification, a category error: it makes video, not meshes.

---

8. WHAT THIS LEAVES OWED (honest gaps in this lane)

• Seed3D 1.0's LICENSE file, at file level, pinned to docs/licence_records/. The single

most load-bearing unverified claim in this report. Zero GPU cost. Do it first.

• The character-benchmark source paper was not pinned. The 100-image / facial-likeness /

VLM-camera-distance protocol described in §1 came through search synthesis; I could not

attribute it to a specific arXiv ID with confidence (the name collides with an unrelated

LLM character-customization benchmark and a 4D animation benchmark). RC5 should pin it

before copying the protocol.

• Rodin/Hyper3D ToS — commercial output rights not stated on the site; unread.
• Sparc3D weights status — my source contradicts the prior survey; unresolved.
• No likeness number exists for TripoSG, our incumbent. Nothing external benchmarks it

on characters. Any comparison we make has to be run by us.

• Every ranking above is desk research. This lane ran no GPU and rendered no arm; per

§2.1 an A/B is only an A/B when both arms are re-rendered by the instrument doing the

comparing. Nothing here is a measured result.

---

9. THE STRONGEST OBJECTION TO THIS LANE'S OWN RECOMMENDATION

Chasing a better generator may be treating the wrong root cause. RC1 and RC2 say the

chain degraded a good sculpt and grafted a bare cage head onto it; RC3 says the native

texture was never even shown. If any of those hold, then a new generator changes nothing —

we would feed a better mesh into the same quality-destroying assembly and grade the same

turntable again. **The cheapest decisive experiment in this whole post-mortem is still

RC3's** (render TRELLIS.2's native PBR straight, beside the assembled chain), because it

costs no licence read, no purchase and no new dependency, and it discriminates between

"our generator is weak" and "our assembly throws quality away." RC4's candidates only

become the answer *after* RC3 shows the native output is also short.

The counter-argument for doing both in parallel: the two §5 probes (Rodin, Tripo) cost

roughly $20–40 total, need no GPU, no install and no sm_120 fight, and would put a

likeness-leading external result beside our own plate within an afternoon. That is a

genuinely cheap upper bound on what the field can do with our exemplars — and if the paid

tools *also* miss the plates, that reframes the entire post-mortem away from tooling and

toward the head/hair/skin stages, which is a far more valuable finding than a model swap.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root