One perfect render isn't a workflow: the AI image and video generator field guide

The five Juggernaut-XL baseline images were structurally clean. OPEN was legible somewhere in all five, but only one placed a legible sign behind the subject; in the others text migrated to clothing or props, with garbling or duplication in three. Exactly three books appeared in one image. All five subjects looked different. The baseline used 1024×1024 output and 25+20 base/refiner steps, with the default negative prompt “text, watermark.”

In the two-part seed-lock check, unchanged inputs returned a pixel-identical image from ComfyUI’s execution cache. That is evidence of cached reproducibility, not an independent recomputation or a guarantee across machines. Keeping the seed while changing the scene to sitting at a desk typing produced a different-looking woman. Three further runs with anti-garbling negative terms produced one cleaner sign, no correct book counts, and persistent text clutter. The small test does not establish a cause for those errors or a universal limit of SDXL.

Turbo used sd_xl_turbo_1.0_fp16, 512×512, one step, CFG 1 and euler_ancestral. Four warm runs took 590–650 milliseconds; the first took 5.4 seconds including loading. All five were coherent but softer than the larger baseline, with no standalone sign and no fully legible OPEN. The speed comparison includes a smaller output and different sampling, not just a model swap.

FLUX used the stock Flux.1 Schnell FP8 template, flux1-schnell-fp8.safetensors, 1024×1024, four steps, CFG 1, euler/simple and random seeds. Warm times were 14.1–14.2 seconds; the first was 19.98 seconds including loading. All five signs read correctly and no cup carried garbled text, but every image had illegible book titles, every book count failed, and subjects differed. Two images contained a second legible OPEN sign. The tested local FLUX checkpoint was [schnell]; Flux 1 Dev was used in the separate hosted OpenArt test.

One SDXL Canny ControlNet run used controlnet_canny_sdxl_small_fp16 at strength 1.00 over the full denoising range, retaining the baseline prompt and scene. It held pose, framing and layout, but changed the face and retained wrong book counts and a duplicated cup; text bleed spread. No IPAdapter or comparable identity tool was installed. This was not a new-pose test. A separate review of 14 existing same-scenario local images found no fully clean source; that review pool is not an additional 14 independent generations to add to the scorecard.

Recraft’s initial Auto batch routed to V4.1 Pro according to each image’s metadata: two 2048×2048 images, 14 credits total, under 20 seconds. Both put the sign off the cup; one sign was cropped. The other image had three books but let OPEN bleed onto two, partly garbled. The subjects shared a general look but were visibly different.

On September 21, the tester used “Add as reference” on the first Recraft image, then requested the same woman sitting at a desk typing in an office. One image, under 30 seconds, cost 19 credits (84 → 65). Hairstyle, earrings, skin tone, facial structure and jacket carried over convincingly, and the pose was followed. This is one reference success, not a reliable identity rate; the original sign/book test did not apply to the new scene. An unrelated animation request already in the project history was rejected for requiring a paid subscription. Its origin and date remain unknown, so it is not counted as a new controlled video attempt.

The free account used Firefly Image 3, Fast mode, square output and default Auto content type, not the offered Image 5 preview. Each click returned four images. The first batch completed in under ten seconds and the repeat in about eight. All eight were visually clean at the gross-artifact level.

The first batch had one legible OPEN on a mug, but no image combined clean sign placement and legibility. The repeat produced two separate sign-like objects, but one was obscured and the other had a malformed final letter: no fully legible OPEN anywhere. Both batches missed every book count. Four distinct-looking subjects became three similar-looking women in the repeat batch with the same settings. That unexplained swing does not prove a durable character-consistency feature. Credits used were not quantified in this record.

The Personal/Free account used Auto, 1:1, four images requested and no style, character or color presets. The first batch delivered four images in under ten seconds despite a slow-queue message; the balance fell 12 → 4, or eight credits total. The identical repeat delivered eight images rather than four, took well over 100 seconds without precise stopwatch timing, and consumed the remaining four credits. That is two credits per image followed by 0.5 credits per image, not a stable rate to budget against.

All 12 outputs had readable standalone signs and blank cups, and subjects appeared consistent within each batch. All 12 missed the three-book instruction. A new-pose reference test was not performed. Neither the changed batch size nor the changed cost has a confirmed explanation; Auto’s underlying model routing was not established.

On Free, the default Nano Banana Pro quote was 40 credits against an eight-credit balance. It was not submitted at that price. The tester substituted Flux 1 Dev at five credits, retaining 4:3 and 1K settings. One clean image arrived in under eight seconds with a legible standalone sign, roughly seven books and garbled secondary text. A large OpenArt watermark and the interface’s export-watermark disclosure limited the free result. Three credits remained.

The same account later displayed Starter and thousands of credits for an unknown reason. Nano Banana Pro then produced one image at the unchanged 40-credit quote, confirmed deducted, in under 20 seconds. The sign was clean; three books were countable, with two legible titles and a third partly visible. No watermark was visible. This is an important exception to the earlier book-count failures, but model, plan and balance all changed. Neither run tested repeatability or subject consistency, and the account change remains unexplained.

Two LTXV-2B distilled text-to-video runs used 30 steps and 768×512. Each produced about four seconds from the template’s 97 frames at 24fps, despite five seconds in the prompt. Grip looked plausible and the scene was coherent, but both clips set down an already-held cup and kept the camera static. Framing prevented a face-identity assessment.

Wan2.2 5B produced one roughly five-second 1280×704 clip at 20 steps. It also appeared to reverse the action and ignored the push-in, while rendering a black leather jacket rather than red. It completed successfully in 27 minutes 28 seconds. System RAM fell below 500 MB free out of 32 GB during an extended decode/encode stage. Memory pressure was observed; precise VRAM use and the causal contribution of each stage were not instrumented. This is a hardware-fit warning for this setup, not a claim about all Wan or local video workflows.

The LTX image-to-video variation used a flawed SDXL source resized from 1024×1024 to 512×512, 30 steps and an LTXVImgToVideo node. In 45.8 seconds it produced about four seconds of a plausible lift-and-sip with a push-in and recognizable source composition. Garbled source text persisted; the final close-up obscured part of the face. Different prompt, conditioning and size mean it cannot isolate which change produced the better action result.

Two paid-account Agent requests used the exact video prompt plus an instruction not to embellish it, with Veo 3.1 - Fast manually selected. The first cost 15 credits and was recorded as a ten-second clip; its generation time and measured pixel dimensions were not recorded. Visual review found a convincing grip, push-in and stable appearance in the checked frames.

The repeat explicitly refused five seconds, claiming the selected model supported only eight seconds on that account’s tier. The eight-second alternative was accepted, delivered and confirmed in the timeline. It cost 20 credits and took about 109 seconds including a high-demand queue message. Grip and push-in repeated. These are observed Agent behaviors, not proof of a universal Veo duration limit. Current documentation lists 4/6/8-second Fast generations, so the tier explanation and first ten-second result remain unverified. No cause is assigned to the 15 → 20 credit difference.

The account had no explicit Free label, but showed upgrade prompts and no active-subscription indicator. Video 3.0 defaults were 1080p, five seconds, 16:9, one clip and Native Audio on. The first run spent 60 of 66 credits. Its interface advertised a one-minute estimate and a Fast-Track message claiming two hours saved; actual observed completion took about 19–20 minutes. The unexperienced two-hour queue is a vendor claim, not measured waiting time.

Before the repeat, the balance rose from six to 336 credits without a recorded purchase or plan change. The second run also cost 60 credits, had no fast-track or wait estimate, and completed in under three minutes. Both media readings were 5.153 seconds. Checked frames showed the cup lifted from the table and a continuous push-in; the face was cropped out after about 2.6 seconds. Nonzero decoded audio bytes supported audio-track presence in both, but listening quality and synchronization were not evaluated. The balance increase and queue differences remain unresolved.

The fresh Account 2 trial succeeded through chat after the direct Create Video node stayed disabled. Its first clip measured 5.041 seconds, used a reported 720p setting, took under two minutes and cost 158 credits. The usage/balance record showed 2,842 trial credits remaining, consistent with an approximately 3,000-credit starting allowance. Checked frames showed lift, grip and push-in without visible warping. The face was not sufficiently visible for identity evaluation. The record reports silence and a visible Luma AI watermark.

The first repeat used the tester’s paraphrase through chat: under two minutes, 178 credits, measured 1280×720 and 5.041 seconds, silent and watermarked, with convincing grip and push-in. The direct node now explicitly displayed “Requires Keyframe(s)”; Create Scene offered a ten-second minimum. The original entry already recorded a disabled direct node, so this evidence establishes a later explicit explanation, not a proven regression from a previously working direct text-to-video path.

Before repeat two, the tester reported personally upgrading to ensure enough credits. The interface continued to display Luma Trial; the reason for that label is unknown. The exact test wording through chat yielded another 5.041-second, 1280×720, silent, watermarked clip with the same checked action and camera behavior. It took about four minutes and cost 559 credits, about 3.1 times the preceding 178-credit run. The agent claimed 1920×1080/1080p, contradicting the delivered videoWidth/videoHeight measurements. The prompt and account funding conditions were not identical across repeats. Neither the charge difference nor the incorrect resolution report has a confirmed cause.

The two-account plan conflict below remains unresolved. These repeats used Account 2 again, not a third independent account.