---
type: "post"
id: "hy-nomas-interactive-demo-gains-image-to-video-5ca1be"
slug: "hy-nomas-interactive-demo-gains-image-to-video-5ca1be"
title: "Hy-Nomas interactive demo gains image-to-video"
description: "Tencent’s Hy-Nomas demo can now initialize an interactive video world from an uploaded image before accepting text prompts and camera control."
retrieval_nugget: "Tencent’s Hy-Nomas demo can now initialize an interactive video world from an uploaded image before accepting text prompts and camera control. For teams evaluating generative video, the change creates a concrete image-to-continuation test path: use a visual reference to establish a scene, then inspect whether later generation preserves it."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["prompts"]
entities: ["Tencent","Tencent Hunyuan","Hy-Nomas","Nomas-PE","DCAE-V","Kimi-K3"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "The Hy-Nomas interactive demo now supports image-to-video generation."
prior_refs: ["https://newruntime.com/posts/tencent-releases-hy-nomas-an-open-interactive-video-c43eeb/"]
source_url: "https://github.com/Tencent-Hunyuan/Hy-Nomas/commit/a31c0fe0bcc493be188fd37ddba4ce2f674becb4"
source_urls: ["https://github.com/Tencent-Hunyuan/Hy-Nomas/commit/a31c0fe0bcc493be188fd37ddba4ce2f674becb4"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:hy-nomas-interactive-demo-gains-image-to-video-5ca1be"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/hy-nomas-interactive-demo-gains-image-to-video-5ca1be/","markdown":"https://newruntime.com/posts/hy-nomas-interactive-demo-gains-image-to-video-5ca1be.md","json":"https://newruntime.com/posts/hy-nomas-interactive-demo-gains-image-to-video-5ca1be.json"}
markdown_url: "https://newruntime.com/posts/hy-nomas-interactive-demo-gains-image-to-video-5ca1be.md"
---

# Hy-Nomas interactive demo gains image-to-video

## Retrieval answer

Tencent’s Hy-Nomas demo can now initialize an interactive video world from an uploaded image before accepting text prompts and camera control. For teams evaluating generative video, the change creates a concrete image-to-continuation test path: use a visual reference to establish a scene, then inspect whether later generation preserves it.

## Summary
Tencent’s Hy-Nomas demo can now initialize an interactive video world from an uploaded image before accepting text prompts and camera control. For teams evaluating generative video, the change creates a concrete image-to-continuation test path: use a visual reference to establish a scene, then inspect whether later generation preserves it.

## What happened
Tencent Hunyuan has added image-reference initialization to the interactive demo for Hy-Nomas. The project’s README announces the feature as “Image reference in the demo”: users upload an image to initialize a world, then continue through text prompts and camera control. Tencent links both the hosted interactive demo and deployment instructions from the repository update.

The deployment guide describes the user flow in detail. Image mode accepts one JPEG, PNG, or WebP file up to 10 MB. Selecting an image automatically switches the interface into image-reference mode, where the user can still choose a perspective but replaces the usual subject and environment fields with a single prompt describing what should happen next. Nomas-PE captions the uploaded image in the background. After captioning completes, the user selects Enter world; the demo generates an initial image-reference segment before it sends the user instruction and that segment’s video history to the selected prompt-enhancement service for segment two.

Tencent’s guide says the reference is internally prepared as a five-second sequence at 24 frames per second: 120 pixel frames become 30 latent frames for the model pipeline. The original image is resized proportionally into a 1280×720 frame and receives centered black padding when its aspect ratio differs. The stated intention is to retain the complete image rather than crop or stretch it. At the default setting, DCAE-V operates at 1280×704 internally because its spatial compression requires dimensions divisible by 32.

The whole five-second reference sequence is retained in the model cache, but it is not all displayed in the downloadable or playable result. The visible T0 reference-image segment consists of its last 0.5 seconds, or 12 frames. The next segment, labeled as text-driven, begins at playback time 0.5 seconds. Tencent says the system does not start the second request until the reference segment succeeds. This makes the image a state-setting input to a multi-segment workflow rather than merely an additional prompt attachment.

The interface includes reset behavior designed around that workflow. Removing or cancelling an image returns the user to text-only mode and restores manual drafts. Starting a new world or refreshing clears the previous image, its caption, and world prompts, while preserving perspective, chosen model, thinking level, inference steps, and seed settings. The guide says those actions also cancel generation or rollback polling for the prior session and request that the old session be released, so late responses cannot overwrite the newly opened world.

For self-hosted use, Tencent documents configuration details as well as the interaction. The gateway and workers require a DCAE-V encoder checkpoint for this path. Image captioning uses a Nomas-PE endpoint configured through environment variables; Kimi-K3 can optionally appear as a prompt-enhancer choice for text and camera prompts, but Tencent says image uploads use Nomas-PE captioning. The repository also includes a new image-reference guide and tests covering the reference-image behavior.

## What the archive adds
New Runtime covered Hy-Nomas as an open interactive video world model on [October 24, 2026](https://newruntime.com/posts/tencent-releases-hy-nomas-an-open-interactive-video-c43eeb/), its SGLang support on [October 28, 2026](https://newruntime.com/posts/sglang-merges-hy-nomas-support-for-offline-and-372c96/), and its model and benchmark publication on [October 31, 2026](https://newruntime.com/posts/tencent-hunyuan-publishes-hy-nomas-an-open-8b-6718e9/). This is a new capability in a known story, not a restatement of the original release.

## Trend signal
This modestly supports [generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/). Hy-Nomas turns image-conditioned generation into a defined application flow with file constraints, preprocessing, cached state, segment sequencing, and reset behavior. That is more operational than requesting a single clip from a standalone model. Still, this is one project’s demo implementation, not proof that the interaction pattern has become a standard product boundary.

## Why it matters
A reference image can specify composition, subject, environment, and visual style more precisely than a text prompt alone. The useful engineering distinction in this change is the handoff: Hy-Nomas first creates a reference-backed video state, then applies the instruction for what happens next using that state’s video history. Teams can therefore test two questions separately: whether the system represents the reference faithfully, and whether its next segment maintains continuity while following a new instruction.

The implementation also offers a useful evaluation discipline for builders. Do not judge the feature from a still frame or the first generated seconds alone. Use a controlled reference image, repeat the same continuation prompt with a text-only setup, and compare subject identity, scene geometry, camera behavior, and the transition at the boundary between T0 and the next segment. Retain seeds and inference settings while varying one input at a time.

This remains a capability for experimenting with Hy-Nomas rather than a general-purpose production workflow. Yet it gives creative-tool teams a more representative test than asking whether a video model can produce an attractive isolated clip.

## Limits and uncertainty
Tencent’s repository describes the feature and implementation, but does not publish comparative quality results for image-initialized versus text-initialized worlds. It also does not report latency, throughput, GPU cost, failure rates, or user-study results. The image flow is documented for the Hy-Nomas demo and its deployment path, so its behavior across different image categories, aspect ratios, long continuations, and self-hosted hardware remains unmeasured here.

The visible output includes only the final 0.5 seconds of the internally generated reference segment, which can make the transition look different from the full initialization process. The repository update also does not independently establish the hosted demo’s operational availability, capacity, or usage limits at any particular time.

## Skill to build now
Build a small continuity-evaluation harness for image-to-video systems. Prepare five reference images covering a person, a product, an interior, a landscape, and a high-contrast graphic. For each, run one identical continuation prompt with fixed seed, perspective, and inference steps; then repeat from text alone. Score reference retention, instruction following, transition quality, and artifacts separately. If the tool exposes a cached reference stage, record whether failures occur before initialization, during the handoff, or in later continuation.

## Reader takeaway
Hy-Nomas now lets an uploaded image establish the first state of an interactive video world before text and camera controls continue it. The material detail is not simply image upload: Tencent sequences a reference segment, retains it in model state, and applies the next instruction only after that segment succeeds. Use it to test continuity, not to assume production reliability.

## Open question
How consistently does the cached reference state preserve identity and scene structure over several subsequent text- or camera-driven segments?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Prompts](https://newruntime.com/topics/prompts.md): Explore the prompts topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
