Essence 2 Max

Official guide to essence-2-max — bitHuman's premium avatar model: the highest-fidelity Essence 2 renderer served on dedicated cloud GPUs, instant identity prep from the internally generated identity video, latency expectations, and pricing.

Note — Generally available. For most agents you simply select essence-2 — the standard model, served the right way for your hardware and quality needs. essence-2-max is the separate, explicitly selectable premium model for maximum-fidelity output — this page documents it. Renamed 2026-07-10 (was essence-2-quality): see Naming & migration.

What it is

essence-2-max (“Essence 2 Max”) is the premium model of the second-generation Essence family — bitHuman’s maximum-fidelity avatar renderer. It is the highest-fidelity Essence 2 renderer served directly: the full rendering stack that the standard Essence 2 is distilled from, animating your identity’s footage on dedicated cloud GPUs and producing hero-quality, close-up output at ~25 frames per second.

Two properties define it:

GPU-only, cloud-only. The renderer needs a server-class GPU; there is no CPU or on-device runtime for this model. (For on-device, CPU, or in-browser serving, use Essence 2 — the standard model serves everywhere.)
No per-identity training. The rendering model is shared; your identity needs only a one-time, lightweight prep that distills the agent’s identity video — generated internally at creation from your portrait image, 10 seconds, authored to loop seamlessly — into a compact identity bundle (a few megabytes). Prep takes seconds, so there is no long training step for Max itself.

When to choose it

Image quality is the whole point — close-up, hero, or brand-critical content where you want the best-looking output bitHuman produces.
True-to-portrait fidelity. Max renders over the identity video generated from your portrait image, so the avatar keeps the person’s true look and framing.
8 credits/min cloud is acceptable — it is the most expensive serving tier, at twice the rate of essence-2 or Expression 2.

If you need scale, CPU serving, or on-device privacy, choose Essence 2. If you want fully generated motion for a stylized character, choose Expression 2. For the family-level decision, start at Essence 2 & Expression 2.

How creation works

Essence 2 Max comes with the combined Essence 2 creation: create the agent with POST /v1/agent/generate and model: "essence-2" — the one asynchronous, 500-credit creation (one-time, per agent) trains the standard Essence 2 and makes Max available from the same internally generated identity video at no extra charge. You pick the model at launch (?model=essence-2-max). You can also add essence-2 to an existing agent that has a stored identity video.

Creation is image-only. Supply a portrait image (or let the prompt generate one) — bitHuman generates the 10-second identity video internally, authored to loop seamlessly (its first and last frames match). Video input is not part of the creation contract and is being removed platform-wide: do not send video — as the rollout completes, a request carrying it is rejected with 400 VIDEO_INPUT_NOT_SUPPORTED before anything is billed.

Like every Essence 2 creation, the input must be a photorealistic human subject — a cartoon / animal / stylized input is rejected up front with 422 MODEL_SUBJECT_MISMATCH, nothing billed (use Expression 2 for those, or model: "auto" to route automatically).

import requests

import os

resp = requests.post(
    "https://api.bithuman.ai/v1/agent/generate",
    headers={
        "Content-Type": "application/json",
        "api-secret": os.environ["BITHUMAN_API_SECRET"],
    },
    json={
        "prompt": "You are a concierge for a luxury hotel.",
        "image": "https://example.com/portrait.jpg",
        "model": "essence-2",   # the combined creation — Max included
    },
)
print(resp.json())
# {"success": true, "message": "Agent generation started",
#  "agent_id": "A66GYD8664", "status": "processing"}

Inputs. Supply an image — a well-lit portrait of the identity facing the camera. bitHuman generates the identity video from it internally (you’ll see current_step: "video" at ~45% progress); Essence 2 Max derives the avatar from that footage, so its look and framing carry into every session. A voice is prepared as part of creation (supply audio to clone one, or one is generated).

What happens. Poll GET /v1/agent/status/{agent_id}: the run moves through the standard steps (payment → persona → voice_image → video), then the model step (reported as current_step: "lip_sync") — Max’s own identity prep distills the compact identity bundle from the generated identity video in seconds. When status reaches ready, the agent is servable.

How long. Max’s identity prep itself takes seconds; the combined creation’s overall wall-clock is set by the standard Essence 2 distillation — typically about 45 minutes (occasionally longer). Keep polling status rather than applying a short client timeout.

Serving

A ready agent serves through every delivery surface — the embed widget, the viewer/share URL, the REST API, and the LiveKit plugin.

Essence 2 Max has a single serving tier:

`?model=` slug	Runtime	Notes
`essence-2-max`	Dedicated cloud GPU (only tier)	Always-warm first line, spilling to elastic cloud GPU overflow that scales from zero.
`essence-2-quality`	(deprecated alias)	The pre-rename slug — the serving worker accepts both, so saved links keep working. Prefer `essence-2-max`. See Naming & migration.

There are no -cpu / -ane runtime slugs for this model, and it is not available on-device — requesting it in an on-device SDK context is reported as cloud-only rather than treated as an unknown model.

Connect behavior. The identity bundle is shared across serving capacity. If a session lands on capacity that hasn’t cached your identity yet, the worker prepares it on the fly (seconds) — you may notice this only as a slightly longer connect on a fresh agent’s first session. The platform also pre-warms overflow capacity while your session connects, so spilling past the first line stays smooth. See session behavior & troubleshooting.

What gates it. The identity prepares on demand, but only from the agent’s stored identity video (generated internally at creation) — so Essence 2 Max is gated on that footage, just as the trained families (Expression 2, the standard Essence 2) are gated on their per-identity models.

An agent with a stored identity video lists the family in its supported_models (status / get / list and the embed-token response — during the rename rollout, responses may report the pre-rename family name; see Naming & migration) and can be requested as this model immediately.
An agent without one cannot render Max and gets a 409 MODEL_NOT_GENERATED rejection naming the missing identity video. To unlock it, create the agent with the combined essence-2 from an image — its pipeline generates the identity video that Max then derives from — or add essence-2 to an agent that already stores one.

Pricing

Surface	Rate
Cloud serving	8 credits/min
Self-hosted serving	4 credits/min
Agent creation	500 credits (one-time — the combined `essence-2` creation; Max included)
Talking-video renders	8 credits per minute of output (rounded up)

Per-minute serving is metered on active avatar minutes only — idle, paused, or disconnected time isn’t billed. Full schedule: Pricing & credits.

Limits and expectations

Renders at ~25 fps on cloud GPUs; video streams over WebRTC.
Image-only creation, via the combined essence-2. Provide a portrait image; the identity video Max derives from is generated internally (10 seconds, authored to loop seamlessly). Video is not a creation input.
Look is fixed at creation. The avatar inherits the generated identity video’s framing and appearance; to change either, create a new agent.
Capacity model: a session takes a dedicated GPU renderer, so Max capacity is deliberately narrower than the standard Essence 2’s — at peak, sessions spill to elastic cloud GPUs that scale from zero, which can add connect time. For high-concurrency deployments, prefer Essence 2.

Next steps

Essence 2 & Expression 2 — the family overview and model chooser.
Second-generation gallery — talk to a live launch agent.
Agents API — full create → poll → serve lifecycle.
Talking video generation — render offline mp4s with essence-2-max.
Session behavior & troubleshooting — latency, idle, common errors.
Download your model — the identity bundle, via API or bithuman pull <code>.
Pricing & credits — plans and the full schedule.