EasyAnimateV5 Sampler for Text to Video

This is the sampler for EasyAnimate's current generation - V5 and V5.1 - and the one the pack defaults to. Give it a model from LoadEasyAnimateModel, a prompt from a text box, and it returns a video as an IMAGE batch. It's text-to-video, but it's also the version of the sampler that shows off what changed in V5: flow matching as a scheduler option and TeaCache, a speed trick, switched on by default.

EasyAnimate V5.1 is Alibaba PAI's answer to the 2025 video-model wave - the 12B is the crowd-pleaser (announced "on par with Hunyuan-Video"), the 7B is the sane consumer choice. It's a DiT with a temporal VAE, Qwen2-VL as the text encoder, and flow-matching sampling: 49 frames max at 8fps, so roughly six seconds. If you want longer, you chain clips - there's no 144-frame mode here the way the old V3/V4 samplers had.

The inputs that matter

• easyanimate_model - from the loader, ideally with V5.1 weights (the plain -zh T2V checkpoints) and the matching config.
• prompt / negative_prompt - STRING_PROMPT, wired from the pack's text boxes. Bilingual: Chinese and English both work.
• width / height - defaults 1008×576, multiples of 16, up to 2048. Since T2V has no image to steal an aspect ratio from, you're in charge.
• video_length - 1 to 49, stepped in 4s, default 49.
• scheduler - six options now: the usual Euler/Euler A/DPM++/PNDM/DDIM plus Flow, which is the default and the right choice for V5.1. Flow matching is the straight-line sampling objective the whole modern video stack runs on; if you switch it off and use Euler, you're leaving quality on the table.
• teacache_threshold - 0.0–1.0, default 0.1.
• enable_teacache - default true. TeaCache skips the transformer's computation on diffusion steps that look similar enough, trading a little quality for a large speedup. On a 12B model that's the difference between "waiting forever" and "waiting a while." Leave it on; drop the threshold if you see artifacts.
• is_image - set true for a single-frame T2I render from the same model.
• seed, steps, cfg - 25 steps, CFG 7 by default. The usual.

The output

images - an IMAGE tensor of frames. Feed it to VideoHelperSuite to save a file.

Where this pack trips people up

The 12B is VRAM-hungry - this is the node where people actually hit OOM. On 24GB, if model_cpu_offload won't do it, flip GPU_memory_mode to model_cpu_offload_and_qfloat8 in the loader; sequential offload is the last resort and it's slow. And expect a chunky system-RAM spike while the model loads. If all you want is to render tonight, the 7B with qfloat8 is the setup that finishes. Motion-wise, set expectations: EasyAnimate's community verdict even at V5.1 was "solid but conservative movement" - it won't blow you away with dynamism, but for clean, controllable 6-second clips it's a genuinely pleasant model to work with.